As a side-effect of running the SpaceTally code I had to look at implementations of #roundTo: because the output was bizarrely formatted.
It turns out that for several of the percentage values calculated by - percent := s spaceForInstances*100.0/totalInstSpace roundTo: 0.1. - we get decidedly not numbers that match what we probably think we should get. For example 28846801 *100.0 / 53172599 -> 54.25125260474855 BUT 54.25125260474855 roundTo: 0.1 -> 54.300000000000004 What? That's not even correct, let alone rounded to the requested precision.
Who is a numerics aficionado?
tim -- tim Rowledge; tim@rowledge.org; http://www.rowledge.org/tim Managing programmers is like herding cats.
tim Rowledge wrote
As a side-effect of running the SpaceTally code I had to look at implementations of #roundTo: because the output was bizarrely formatted.
It turns out that for several of the percentage values calculated by - percent := s spaceForInstances*100.0/totalInstSpace roundTo: 0.1.
- we get decidedly not numbers that match what we probably think we should
get. For example 28846801 *100.0 / 53172599 -> 54.25125260474855 BUT 54.25125260474855 roundTo: 0.1 -> 54.300000000000004 What? That's not even correct, let alone rounded to the requested precision.
Who is a numerics aficionado?
tim
tim Rowledge;
tim@
; http://www.rowledge.org/tim Managing programmers is like herding cats.
roundTo: rounds to a quantum so:
100 roundTo: 8 -> 104 100 roundTo: 0.8 -> 100 100 roundTo: (Float pi) -> 100.53096491487338
For printing you want: 54.25125260474855 printShowingDecimalPlaces: 1 -> '54.3'
-- View this message in context: http://forum.world.st/The-joys-or-not-of-floating-point-numbers-tp4678288p46... Sent from the Squeak - Dev mailing list archive at Nabble.com.
2013/3/26 glenpaling slp5591@me.com:
tim Rowledge wrote
As a side-effect of running the SpaceTally code I had to look at implementations of #roundTo: because the output was bizarrely formatted.
It turns out that for several of the percentage values calculated by - percent := s spaceForInstances*100.0/totalInstSpace roundTo: 0.1.
- we get decidedly not numbers that match what we probably think we should
get. For example 28846801 *100.0 / 53172599 -> 54.25125260474855 BUT 54.25125260474855 roundTo: 0.1 -> 54.300000000000004 What? That's not even correct, let alone rounded to the requested precision.
Who is a numerics aficionado?
tim
tim Rowledge;
tim@
; http://www.rowledge.org/tim Managing programmers is like herding cats.
roundTo: rounds to a quantum so:
100 roundTo: 8 -> 104 100 roundTo: 0.8 -> 100 100 roundTo: (Float pi) -> 100.53096491487338
For printing you want: 54.25125260474855 printShowingDecimalPlaces: 1 -> '54.3'
Yep, it's because we now tell the awfull truth about floats.
0.1 = (1/10) -> false. "it's not exactly 1/10" 0.1 = (1/10.0) -> true. "yes, it's the same approximation, with same rounding error"
#(successor predecessor) allSatisfy: [:neighbourhood | ((0.1 perform: neighbourhood) asFraction - (1/10)) abs >= (0.1 asFraction - (1/10)) abs]. "yes 0.1 is closest float to 1/10"
(0.1 asFraction - (1/10)) / 0.1 ulp -> 0.4. "Yep, 0.4 ulp error is OK, IEEE 754 ops guaranty +/- 0.5 ulp"
but: 543/10.0 -> 54.3. "sounds good" 543/10.0 = (543/10) -> false. "But inexact, with a single rounding error" 543*0.1-> 54.300000000000004. "this cumulated two rounding errors"
(54.3 asFraction- (543/10)) / 54.3 ulp -> -0.4. ((543*0.1) asFraction- (543/10)) / 54.3 ulp -> 0.6.
So what you want is 543/10, but that's not what roundTo: 0.1 does...
Old 3.X Squeak could maintain the illusion a little longer because it used to print Floats approximately. Float>>printString now prints the shortest decimal form that will be re-interpreted as the same Float, and that's the only thing that changed. Every two finite different Float shall have a different print.
Use printShowingDecimalPlaces: as suggested or round to an exact fraction
54.25125260474855 roundTo: 1/10. 54.25125260474855 roundTo: 0.1s1.
Nicolas
-- View this message in context: http://forum.world.st/The-joys-or-not-of-floating-point-numbers-tp4678288p46... Sent from the Squeak - Dev mailing list archive at Nabble.com.
In my image I have:
precision: aNumber
| pten |
self isInfinite ifTrue: [^ self].
pten := 10 raisedTo: aNumber. ^ (pten * self) rounded / pten asFloat
54.25125260474855 precision: 1 ==> 54.3
What do you think of it Nicolas ? Does it make sense ?
Stef
2013/3/26 Stéphane Rollandin lecteur@zogotounga.net:
In my image I have:
precision: aNumber
| pten | self isInfinite ifTrue: [^ self]. pten := 10 raisedTo: aNumber. ^ (pten * self) rounded / pten asFloat
54.25125260474855 precision: 1 ==> 54.3
What do you think of it Nicolas ? Does it make sense ?
Stef
Yes, it's better than roundTo: 0.1. Pharo has implemented a similar method named #round: http://code.google.com/p/pharo/issues/detail?id=5590 It's still not perfect because it cumulates two inexact operations: - one inexact operation in the multiplication pten * self - one inexact operation in the division / pten
Example (0.995 round: 2) -> 1.00 though 0.995 < (995/1000) so it should round to 0.99.
Nicolas
It's still not perfect because it cumulates two inexact operations:
- one inexact operation in the multiplication pten * self
- one inexact operation in the division / pten
Example (0.995 round: 2) -> 1.00 though 0.995 < (995/1000) so it should round to 0.99.
Thanks. This stuff is really tricky !
Stef
On Tue, 26 Mar 2013 18:32:31 +0100, Stéphane Rollandin lecteur@zogotounga.net wrote:
Example (0.995 round: 2) -> 1.00 though 0.995 < (995/1000) so it should round to 0.99.
.. and I just noticed that
0.995 printShowingDecimalPlaces: 2 ==> '1.00'
Do we have a method returning the correct value ?
That looks right to me, because 0.995 was rounded up.
Lou ----------------------------------------------------------- Louis LaBrunda Keystone Software Corp. SkypeMe callto://PhotonDemon mailto:Lou@Keystone-Software.com http://www.Keystone-Software.com
2013/3/26 Louis LaBrunda Lou@keystone-software.com:
On Tue, 26 Mar 2013 18:32:31 +0100, Stéphane Rollandin lecteur@zogotounga.net wrote:
Example (0.995 round: 2) -> 1.00 though 0.995 < (995/1000) so it should round to 0.99.
.. and I just noticed that
0.995 printShowingDecimalPlaces: 2 ==> '1.00'
Do we have a method returning the correct value ?
That looks right to me, because 0.995 was rounded up.
Lou
No it's not, because 0.995 < 0.995s3 so it should be rounded down. In a recent trunk image, print is OK: (0.995 printShowingDecimalPlaces: 2) -> '0.99'
Nicolas
Louis LaBrunda Keystone Software Corp. SkypeMe callto://PhotonDemon mailto:Lou@Keystone-Software.com http://www.Keystone-Software.com
On Tue, 26 Mar 2013 18:59:12 +0100, Nicolas Cellier nicolas.cellier.aka.nice@gmail.com wrote:
2013/3/26 Louis LaBrunda Lou@keystone-software.com:
On Tue, 26 Mar 2013 18:32:31 +0100, Stéphane Rollandin lecteur@zogotounga.net wrote:
Example (0.995 round: 2) -> 1.00 though 0.995 < (995/1000) so it should round to 0.99.
.. and I just noticed that
0.995 printShowingDecimalPlaces: 2 ==> '1.00'
Do we have a method returning the correct value ?
That looks right to me, because 0.995 was rounded up.
Lou
No it's not, because 0.995 < 0.995s3 so it should be rounded down. In a recent trunk image, print is OK: (0.995 printShowingDecimalPlaces: 2) -> '0.99'
In VA Smalltalk:
0.995 = 0.995s3 => true 0.995 < 0.995s3 => false
I think in VA Smalltalk there is some VM magic going on that makes the above work the way it does. The #= and #< of Float are implemented in a primitive. I guess when converting the '0.995' string to a float a little is lost and that would make it less than 0.995s3 but there is a lot of code floating around (sorry about the puns) to make floats look better. In that case I would think people would want (0.995 printShowingDecimalPlaces: 2) to show '1.00' and not '0.99'.
Anyway, we should try not to use floats without a very, very good reason. Like we have to send them outside of Smalltalk or we really need the speed or decimals and fractions take up too much memory. But then we must live with their inaccuracies and display mess.
I have an untested theory that fractions can be close in speed to floats because divisions (that are expensive) can be pushed to the end of a computation because with fractions they are multiplies.
Lou ----------------------------------------------------------- Louis LaBrunda Keystone Software Corp. SkypeMe callto://PhotonDemon mailto:Lou@Keystone-Software.com http://www.Keystone-Software.com
2013/3/26 Louis LaBrunda Lou@keystone-software.com:
On Tue, 26 Mar 2013 18:59:12 +0100, Nicolas Cellier nicolas.cellier.aka.nice@gmail.com wrote:
2013/3/26 Louis LaBrunda Lou@keystone-software.com:
On Tue, 26 Mar 2013 18:32:31 +0100, Stéphane Rollandin lecteur@zogotounga.net wrote:
Example (0.995 round: 2) -> 1.00 though 0.995 < (995/1000) so it should round to 0.99.
.. and I just noticed that
0.995 printShowingDecimalPlaces: 2 ==> '1.00'
Do we have a method returning the correct value ?
That looks right to me, because 0.995 was rounded up.
Lou
No it's not, because 0.995 < 0.995s3 so it should be rounded down. In a recent trunk image, print is OK: (0.995 printShowingDecimalPlaces: 2) -> '0.99'
In VA Smalltalk:
0.995 = 0.995s3 => true 0.995 < 0.995s3 => false
This is a default st80 behavior which converts minimumGenerality number to maximumGenerality. Maybe some of these conventions are hardwired in VM, I don't know, but in other dialects it's handled at image side.
The idea was this one: if you perform an operation between an inexact value and an exact value, the result is inexact. So Float + * / - Integer,Fraction,ScaledDecimal will result into a Float. Thus Float got a higher generality.
The idea behind Squeak change is that every Float has an exact value (asTrueFraction). Since we have only two possible answers true/false for comparison, and no maybe or dontKnow, it makes sense to compare the exact value. This reduces the number of paradoxal equalities
| a b c | a := 1 << Float precision. b := a + 1. c := a asFloat. { c = b. c = a. a = b. }
In VW and VA, the first two are true, the last is false, which suggests that = is not an equivalence relation. In Squeak, only the second one is true.
Same with inequalities, we expect (a < b) & (a = c) ==> (c < b) etc... This is still true in Squeak, not in VA/VW/st80. I think gst and Dolphin and maybe stx adopted Squeak behavior (to be confirmed).
I think in VA Smalltalk there is some VM magic going on that makes the above work the way it does. The #= and #< of Float are implemented in a primitive. I guess when converting the '0.995' string to a float a little is lost and that would make it less than 0.995s3 but there is a lot of code floating around (sorry about the puns) to make floats look better. In that case I would think people would want (0.995 printShowingDecimalPlaces: 2) to show '1.00' and not '0.99'.
No, people should not rely on such behavior with Floats, because sooner than later they will be bitten. We cannot cheat very long with this kind of assumptions. My POV is that it is better to educate about false expectations with Float, and that's the main benefit of (1/10) ~= 0.1.
I would add that such print policy is not the behaviour of every other language I know of.
printf( "%.2f",0.995) -> 0.99
Because libm are carefully written nowdays and other languages libraries are either built over libm or much more careful than Smalltalk were (that mostly means more recent).
Anyway, we should try not to use floats without a very, very good reason. Like we have to send them outside of Smalltalk or we really need the speed or decimals and fractions take up too much memory. But then we must live with their inaccuracies and display mess.
Good advice, let's put those expectations on decimal fractions (ScaledDecimal/FixedPoint) or general Fraction.
I have an untested theory that fractions can be close in speed to floats because divisions (that are expensive) can be pushed to the end of a computation because with fractions they are multiplies.
Lou
Why not, but huge numerators and denominators are not cheap compared to Float operations. OK reducing the fraction is the expensive part, but how does the cost grow versus the length of operands?
Also some geometric operations are not even possible on Q (hypot) so you might soon need AlgebraicNumber. And Smalltalk is also about graphics and geometry.
Nicolas
Louis LaBrunda Keystone Software Corp. SkypeMe callto://PhotonDemon mailto:Lou@Keystone-Software.com http://www.Keystone-Software.com
Hi Nicolas,
snip...
In VA Smalltalk:
0.995 = 0.995s3 => true 0.995 < 0.995s3 => false
This is a default st80 behavior which converts minimumGenerality number to maximumGenerality.
In VA when one converts 0.995 to decimal, one gets 0.995s15, which when compared to 0.995s3 is equal. And when 0.995s3 is converted to float that float compares equal to 0.995.
Maybe some of these conventions are hardwired in VM, I don't know, but in other dialects it's handled at image side.
The idea was this one: if you perform an operation between an inexact value and an exact value, the result is inexact. So Float + * / - Integer,Fraction,ScaledDecimal will result into a Float. Thus Float got a higher generality.
This would imply that one should convert the ScaledDecimal to a Float before the compare. But as above that would not change things (I think maybe because VA uses 8 byte floats).
This is from Squeak:
0.995s3 asFloat 0.995
0.995s3 asFloat = 0.995 true
0.995 asScaledDecimal 0.99499999s8
(995 / 1000) asScaledDecimal 0.995s3
0.995 asTrueFraction asScaledDecimal 0.99499999999999999555910790149937383830547332763671875s53
(0.995 * 1000000000) asScaledDecimal / 1000000000 0.99500000s8
0.995 * 1000000000 9.95e8
Sorry, but I have run out of time to play at the moment. So, I will just throw a thought out there. I think there may be a problem with asTrueFraction. Which if implemented differently might not make 0.995 < 0.995s3.
The idea behind Squeak change is that every Float has an exact value (asTrueFraction). Since we have only two possible answers true/false for comparison, and no maybe or dontKnow, it makes sense to compare the exact value. This reduces the number of paradoxal equalities
| a b c | a := 1 << Float precision. b := a + 1. c := a asFloat. { c = b. c = a. a = b. }
In VW and VA, the first two are true, the last is false, which suggests that = is not an equivalence relation. In Squeak, only the second one is true.
Same with inequalities, we expect (a < b) & (a = c) ==> (c < b) etc... This is still true in Squeak, not in VA/VW/st80. I think gst and Dolphin and maybe stx adopted Squeak behavior (to be confirmed).
I think in VA Smalltalk there is some VM magic going on that makes the above work the way it does. The #= and #< of Float are implemented in a primitive. I guess when converting the '0.995' string to a float a little is lost and that would make it less than 0.995s3 but there is a lot of code floating around (sorry about the puns) to make floats look better. In that case I would think people would want (0.995 printShowingDecimalPlaces: 2) to show '1.00' and not '0.99'.
No, people should not rely on such behavior with Floats, because sooner than later they will be bitten. We cannot cheat very long with this kind of assumptions. My POV is that it is better to educate about false expectations with Float, and that's the main benefit of (1/10) ~= 0.1.
I would add that such print policy is not the behaviour of every other language I know of.
printf( "%.2f",0.995) -> 0.99
Because libm are carefully written nowdays and other languages libraries are either built over libm or much more careful than Smalltalk were (that mostly means more recent).
Anyway, we should try not to use floats without a very, very good reason. Like we have to send them outside of Smalltalk or we really need the speed or decimals and fractions take up too much memory. But then we must live with their inaccuracies and display mess.
Good advice, let's put those expectations on decimal fractions (ScaledDecimal/FixedPoint) or general Fraction.
I have an untested theory that fractions can be close in speed to floats because divisions (that are expensive) can be pushed to the end of a computation because with fractions they are multiplies.
Lou
Why not, but huge numerators and denominators are not cheap compared to Float operations. OK reducing the fraction is the expensive part, but how does the cost grow versus the length of operands?
Also some geometric operations are not even possible on Q (hypot) so you might soon need AlgebraicNumber. And Smalltalk is also about graphics and geometry.
Nicolas
Louis LaBrunda Keystone Software Corp. SkypeMe callto://PhotonDemon mailto:Lou@Keystone-Software.com http://www.Keystone-Software.com
----------------------------------------------------------- Louis LaBrunda Keystone Software Corp. SkypeMe callto://PhotonDemon mailto:Lou@Keystone-Software.com http://www.Keystone-Software.com
2013/3/26 Louis LaBrunda Lou@keystone-software.com:
Hi Nicolas,
snip...
In VA Smalltalk:
0.995 = 0.995s3 => true 0.995 < 0.995s3 => false
This is a default st80 behavior which converts minimumGenerality number to maximumGenerality.
In VA when one converts 0.995 to decimal, one gets 0.995s15, which when compared to 0.995s3 is equal. And when 0.995s3 is converted to float that float compares equal to 0.995.
Maybe some of these conventions are hardwired in VM, I don't know, but in other dialects it's handled at image side.
The idea was this one: if you perform an operation between an inexact value and an exact value, the result is inexact. So Float + * / - Integer,Fraction,ScaledDecimal will result into a Float. Thus Float got a higher generality.
This would imply that one should convert the ScaledDecimal to a Float before the compare. But as above that would not change things (I think maybe because VA uses 8 byte floats).
This is from Squeak:
0.995s3 asFloat 0.995
0.995s3 asFloat = 0.995 true
0.995 asScaledDecimal 0.99499999s8
(995 / 1000) asScaledDecimal 0.995s3
0.995 asTrueFraction asScaledDecimal 0.99499999999999999555910790149937383830547332763671875s53
(0.995 * 1000000000) asScaledDecimal / 1000000000 0.99500000s8
Yep, but (0.995 * 1000000000) is inexact... You can check that (0.995 * 1000000000) ~~ (0.995 asFraction * 1000000000)
You cannot rely on results of any such inexact calculus, or you'll be building on sand.
0.995 * 1000000000 9.95e8
Sorry, but I have run out of time to play at the moment. So, I will just throw a thought out there. I think there may be a problem with asTrueFraction. Which if implemented differently might not make 0.995 < 0.995s3.
Well I doubt, but I'm all ears ;)
(0.995 asFraction storeStringBase: 2) -> '(2r11111110101110000101000111101011100001010001111010111/2r100000000000000000000000000000000000000000000000000000)'.
(0.995 successor) asFraction storeStringBase: 2 -> '(2r11111110101110000101000111101011100001010001111011/2r100000000000000000000000000000000000000000000000000)' -> '(2r11111110101110000101000111101011100001010001111011000/2r100000000000000000000000000000000000000000000000000000)'.
(0.995 predecessor) asFraction storeStringBase: 2 -> '(2r1111111010111000010100011110101110000101000111101011/2r10000000000000000000000000000000000000000000000000000)' -> '(2r11111110101110000101000111101011100001010001111010110/2r100000000000000000000000000000000000000000000000000000)'.
0.995 ulp asFraction storeStringBase: 2 -> '(2r1/2r100000000000000000000000000000000000000000000000000000)'.
(995/1000 - 0.995 asFraction) / 0.995 ulp -> 0.04.
(995/1000 - 0.995 successor asFraction) / 0.995 ulp -> -0.96.
(995/1000 - 0.995 successor asFraction) / 0.995 ulp -> 1.04.
and numerator/denominator have this property:
2r11111110101110000101000111101011100001010001111010111 highBit -> 53. 2r100000000000000000000000000000000000000000000000000000 highBit -> 54.
So, 53 bits of significand, a power of two on denominator, that sounds like a correct Float, slightly smaller than 1. Next significand and previous significand are both further of (995/1000) than 0.995 is. 0.995 is closest Float to 995/1000 and is smaller than 995/1000, no doubt.
Nicolas
The idea behind Squeak change is that every Float has an exact value (asTrueFraction). Since we have only two possible answers true/false for comparison, and no maybe or dontKnow, it makes sense to compare the exact value. This reduces the number of paradoxal equalities
| a b c | a := 1 << Float precision. b := a + 1. c := a asFloat. { c = b. c = a. a = b. }
In VW and VA, the first two are true, the last is false, which suggests that = is not an equivalence relation. In Squeak, only the second one is true.
Same with inequalities, we expect (a < b) & (a = c) ==> (c < b) etc... This is still true in Squeak, not in VA/VW/st80. I think gst and Dolphin and maybe stx adopted Squeak behavior (to be confirmed).
I think in VA Smalltalk there is some VM magic going on that makes the above work the way it does. The #= and #< of Float are implemented in a primitive. I guess when converting the '0.995' string to a float a little is lost and that would make it less than 0.995s3 but there is a lot of code floating around (sorry about the puns) to make floats look better. In that case I would think people would want (0.995 printShowingDecimalPlaces: 2) to show '1.00' and not '0.99'.
No, people should not rely on such behavior with Floats, because sooner than later they will be bitten. We cannot cheat very long with this kind of assumptions. My POV is that it is better to educate about false expectations with Float, and that's the main benefit of (1/10) ~= 0.1.
I would add that such print policy is not the behaviour of every other language I know of.
printf( "%.2f",0.995) -> 0.99
Because libm are carefully written nowdays and other languages libraries are either built over libm or much more careful than Smalltalk were (that mostly means more recent).
Anyway, we should try not to use floats without a very, very good reason. Like we have to send them outside of Smalltalk or we really need the speed or decimals and fractions take up too much memory. But then we must live with their inaccuracies and display mess.
Good advice, let's put those expectations on decimal fractions (ScaledDecimal/FixedPoint) or general Fraction.
I have an untested theory that fractions can be close in speed to floats because divisions (that are expensive) can be pushed to the end of a computation because with fractions they are multiplies.
Lou
Why not, but huge numerators and denominators are not cheap compared to Float operations. OK reducing the fraction is the expensive part, but how does the cost grow versus the length of operands?
Also some geometric operations are not even possible on Q (hypot) so you might soon need AlgebraicNumber. And Smalltalk is also about graphics and geometry.
Nicolas
Louis LaBrunda Keystone Software Corp. SkypeMe callto://PhotonDemon mailto:Lou@Keystone-Software.com http://www.Keystone-Software.com
Louis LaBrunda Keystone Software Corp. SkypeMe callto://PhotonDemon mailto:Lou@Keystone-Software.com http://www.Keystone-Software.com
2013/3/26 Nicolas Cellier nicolas.cellier.aka.nice@gmail.com:
2013/3/26 Louis LaBrunda Lou@keystone-software.com:
Hi Nicolas,
snip...
In VA Smalltalk:
0.995 = 0.995s3 => true 0.995 < 0.995s3 => false
This is a default st80 behavior which converts minimumGenerality number to maximumGenerality.
In VA when one converts 0.995 to decimal, one gets 0.995s15, which when compared to 0.995s3 is equal. And when 0.995s3 is converted to float that float compares equal to 0.995.
Maybe some of these conventions are hardwired in VM, I don't know, but in other dialects it's handled at image side.
The idea was this one: if you perform an operation between an inexact value and an exact value, the result is inexact. So Float + * / - Integer,Fraction,ScaledDecimal will result into a Float. Thus Float got a higher generality.
This would imply that one should convert the ScaledDecimal to a Float before the compare. But as above that would not change things (I think maybe because VA uses 8 byte floats).
This is from Squeak:
0.995s3 asFloat 0.995
0.995s3 asFloat = 0.995 true
0.995 asScaledDecimal 0.99499999s8
(995 / 1000) asScaledDecimal 0.995s3
0.995 asTrueFraction asScaledDecimal 0.99499999999999999555910790149937383830547332763671875s53
(0.995 * 1000000000) asScaledDecimal / 1000000000 0.99500000s8
Yep, but (0.995 * 1000000000) is inexact... You can check that (0.995 * 1000000000) ~~ (0.995 asFraction * 1000000000)
Oops, I mean (0.995 * 1000000000) ~= (0.995 asFraction * 1000000000)
You cannot rely on results of any such inexact calculus, or you'll be building on sand.
0.995 * 1000000000 9.95e8
Sorry, but I have run out of time to play at the moment. So, I will just throw a thought out there. I think there may be a problem with asTrueFraction. Which if implemented differently might not make 0.995 < 0.995s3.
Well I doubt, but I'm all ears ;)
(0.995 asFraction storeStringBase: 2) -> '(2r11111110101110000101000111101011100001010001111010111/2r100000000000000000000000000000000000000000000000000000)'.
(0.995 successor) asFraction storeStringBase: 2 -> '(2r11111110101110000101000111101011100001010001111011/2r100000000000000000000000000000000000000000000000000)' -> '(2r11111110101110000101000111101011100001010001111011000/2r100000000000000000000000000000000000000000000000000000)'.
(0.995 predecessor) asFraction storeStringBase: 2 -> '(2r1111111010111000010100011110101110000101000111101011/2r10000000000000000000000000000000000000000000000000000)' -> '(2r11111110101110000101000111101011100001010001111010110/2r100000000000000000000000000000000000000000000000000000)'.
0.995 ulp asFraction storeStringBase: 2 -> '(2r1/2r100000000000000000000000000000000000000000000000000000)'.
(995/1000 - 0.995 asFraction) / 0.995 ulp -> 0.04.
(995/1000 - 0.995 successor asFraction) / 0.995 ulp -> -0.96.
(995/1000 - 0.995 successor asFraction) / 0.995 ulp -> 1.04.
and numerator/denominator have this property:
2r11111110101110000101000111101011100001010001111010111 highBit -> 53. 2r100000000000000000000000000000000000000000000000000000 highBit -> 54.
So, 53 bits of significand, a power of two on denominator, that sounds like a correct Float, slightly smaller than 1. Next significand and previous significand are both further of (995/1000) than 0.995 is. 0.995 is closest Float to 995/1000 and is smaller than 995/1000, no doubt.
Nicolas
The idea behind Squeak change is that every Float has an exact value (asTrueFraction). Since we have only two possible answers true/false for comparison, and no maybe or dontKnow, it makes sense to compare the exact value. This reduces the number of paradoxal equalities
| a b c | a := 1 << Float precision. b := a + 1. c := a asFloat. { c = b. c = a. a = b. }
In VW and VA, the first two are true, the last is false, which suggests that = is not an equivalence relation. In Squeak, only the second one is true.
Same with inequalities, we expect (a < b) & (a = c) ==> (c < b) etc... This is still true in Squeak, not in VA/VW/st80. I think gst and Dolphin and maybe stx adopted Squeak behavior (to be confirmed).
I think in VA Smalltalk there is some VM magic going on that makes the above work the way it does. The #= and #< of Float are implemented in a primitive. I guess when converting the '0.995' string to a float a little is lost and that would make it less than 0.995s3 but there is a lot of code floating around (sorry about the puns) to make floats look better. In that case I would think people would want (0.995 printShowingDecimalPlaces: 2) to show '1.00' and not '0.99'.
No, people should not rely on such behavior with Floats, because sooner than later they will be bitten. We cannot cheat very long with this kind of assumptions. My POV is that it is better to educate about false expectations with Float, and that's the main benefit of (1/10) ~= 0.1.
I would add that such print policy is not the behaviour of every other language I know of.
printf( "%.2f",0.995) -> 0.99
Because libm are carefully written nowdays and other languages libraries are either built over libm or much more careful than Smalltalk were (that mostly means more recent).
Anyway, we should try not to use floats without a very, very good reason. Like we have to send them outside of Smalltalk or we really need the speed or decimals and fractions take up too much memory. But then we must live with their inaccuracies and display mess.
Good advice, let's put those expectations on decimal fractions (ScaledDecimal/FixedPoint) or general Fraction.
I have an untested theory that fractions can be close in speed to floats because divisions (that are expensive) can be pushed to the end of a computation because with fractions they are multiplies.
Lou
Why not, but huge numerators and denominators are not cheap compared to Float operations. OK reducing the fraction is the expensive part, but how does the cost grow versus the length of operands?
Also some geometric operations are not even possible on Q (hypot) so you might soon need AlgebraicNumber. And Smalltalk is also about graphics and geometry.
Nicolas
Louis LaBrunda Keystone Software Corp. SkypeMe callto://PhotonDemon mailto:Lou@Keystone-Software.com http://www.Keystone-Software.com
Louis LaBrunda Keystone Software Corp. SkypeMe callto://PhotonDemon mailto:Lou@Keystone-Software.com http://www.Keystone-Software.com
2013/3/26 Nicolas Cellier nicolas.cellier.aka.nice@gmail.com:
2013/3/26 Nicolas Cellier nicolas.cellier.aka.nice@gmail.com:
2013/3/26 Louis LaBrunda Lou@keystone-software.com:
Hi Nicolas,
snip...
In VA Smalltalk:
0.995 = 0.995s3 => true 0.995 < 0.995s3 => false
This is a default st80 behavior which converts minimumGenerality number to maximumGenerality.
In VA when one converts 0.995 to decimal, one gets 0.995s15, which when compared to 0.995s3 is equal. And when 0.995s3 is converted to float that float compares equal to 0.995.
Maybe some of these conventions are hardwired in VM, I don't know, but in other dialects it's handled at image side.
The idea was this one: if you perform an operation between an inexact value and an exact value, the result is inexact. So Float + * / - Integer,Fraction,ScaledDecimal will result into a Float. Thus Float got a higher generality.
This would imply that one should convert the ScaledDecimal to a Float before the compare. But as above that would not change things (I think maybe because VA uses 8 byte floats).
This is from Squeak:
0.995s3 asFloat 0.995
0.995s3 asFloat = 0.995 true
0.995 asScaledDecimal 0.99499999s8
(995 / 1000) asScaledDecimal 0.995s3
0.995 asTrueFraction asScaledDecimal 0.99499999999999999555910790149937383830547332763671875s53
(0.995 * 1000000000) asScaledDecimal / 1000000000 0.99500000s8
Yep, but (0.995 * 1000000000) is inexact... You can check that (0.995 * 1000000000) ~~ (0.995 asFraction * 1000000000)
Oops, I mean (0.995 * 1000000000) ~= (0.995 asFraction * 1000000000)
Another way to say it is:
(0.995 asFraction * 1000000000) numerator highBit > Float precision.
As denominator is a power of two and resulting fraction is reduced, numerator is odd (or zero). So the numerator highBit is the number of bits of the significand of exact result. So with more bits (binary digits) than a Float significand can contain, Float result can't be exact.
Nicolas
You cannot rely on results of any such inexact calculus, or you'll be building on sand.
0.995 * 1000000000 9.95e8
Sorry, but I have run out of time to play at the moment. So, I will just throw a thought out there. I think there may be a problem with asTrueFraction. Which if implemented differently might not make 0.995 < 0.995s3.
Well I doubt, but I'm all ears ;)
(0.995 asFraction storeStringBase: 2) -> '(2r11111110101110000101000111101011100001010001111010111/2r100000000000000000000000000000000000000000000000000000)'.
(0.995 successor) asFraction storeStringBase: 2 -> '(2r11111110101110000101000111101011100001010001111011/2r100000000000000000000000000000000000000000000000000)' -> '(2r11111110101110000101000111101011100001010001111011000/2r100000000000000000000000000000000000000000000000000000)'.
(0.995 predecessor) asFraction storeStringBase: 2 -> '(2r1111111010111000010100011110101110000101000111101011/2r10000000000000000000000000000000000000000000000000000)' -> '(2r11111110101110000101000111101011100001010001111010110/2r100000000000000000000000000000000000000000000000000000)'.
0.995 ulp asFraction storeStringBase: 2 -> '(2r1/2r100000000000000000000000000000000000000000000000000000)'.
(995/1000 - 0.995 asFraction) / 0.995 ulp -> 0.04.
(995/1000 - 0.995 successor asFraction) / 0.995 ulp -> -0.96.
(995/1000 - 0.995 successor asFraction) / 0.995 ulp -> 1.04.
and numerator/denominator have this property:
2r11111110101110000101000111101011100001010001111010111 highBit -> 53. 2r100000000000000000000000000000000000000000000000000000 highBit -> 54.
So, 53 bits of significand, a power of two on denominator, that sounds like a correct Float, slightly smaller than 1. Next significand and previous significand are both further of (995/1000) than 0.995 is. 0.995 is closest Float to 995/1000 and is smaller than 995/1000, no doubt.
Nicolas
The idea behind Squeak change is that every Float has an exact value (asTrueFraction). Since we have only two possible answers true/false for comparison, and no maybe or dontKnow, it makes sense to compare the exact value. This reduces the number of paradoxal equalities
| a b c | a := 1 << Float precision. b := a + 1. c := a asFloat. { c = b. c = a. a = b. }
In VW and VA, the first two are true, the last is false, which suggests that = is not an equivalence relation. In Squeak, only the second one is true.
Same with inequalities, we expect (a < b) & (a = c) ==> (c < b) etc... This is still true in Squeak, not in VA/VW/st80. I think gst and Dolphin and maybe stx adopted Squeak behavior (to be confirmed).
I think in VA Smalltalk there is some VM magic going on that makes the above work the way it does. The #= and #< of Float are implemented in a primitive. I guess when converting the '0.995' string to a float a little is lost and that would make it less than 0.995s3 but there is a lot of code floating around (sorry about the puns) to make floats look better. In that case I would think people would want (0.995 printShowingDecimalPlaces: 2) to show '1.00' and not '0.99'.
No, people should not rely on such behavior with Floats, because sooner than later they will be bitten. We cannot cheat very long with this kind of assumptions. My POV is that it is better to educate about false expectations with Float, and that's the main benefit of (1/10) ~= 0.1.
I would add that such print policy is not the behaviour of every other language I know of.
printf( "%.2f",0.995) -> 0.99
Because libm are carefully written nowdays and other languages libraries are either built over libm or much more careful than Smalltalk were (that mostly means more recent).
Anyway, we should try not to use floats without a very, very good reason. Like we have to send them outside of Smalltalk or we really need the speed or decimals and fractions take up too much memory. But then we must live with their inaccuracies and display mess.
Good advice, let's put those expectations on decimal fractions (ScaledDecimal/FixedPoint) or general Fraction.
I have an untested theory that fractions can be close in speed to floats because divisions (that are expensive) can be pushed to the end of a computation because with fractions they are multiplies.
Lou
Why not, but huge numerators and denominators are not cheap compared to Float operations. OK reducing the fraction is the expensive part, but how does the cost grow versus the length of operands?
Also some geometric operations are not even possible on Q (hypot) so you might soon need AlgebraicNumber. And Smalltalk is also about graphics and geometry.
Nicolas
Louis LaBrunda Keystone Software Corp. SkypeMe callto://PhotonDemon mailto:Lou@Keystone-Software.com http://www.Keystone-Software.com
Louis LaBrunda Keystone Software Corp. SkypeMe callto://PhotonDemon mailto:Lou@Keystone-Software.com http://www.Keystone-Software.com
Hi Nicolas,
Sorry to be getting back to this so late. I take back what I said about asTrueFraction. I'm sure it is fine. But I think I question its use in this case.
This all started when Stéphane said this:
0.995 printShowingDecimalPlaces: 2 ==> '1.00'
was wrong and I didn't think so.
When one uses a method like printShowingDecimalPlaces: one expects (maybe there is some other similar method that one should expect this) the result to look good (yes that is very subjective) but I would say looking good is more important than being accurate. Yes, that is an admission that asTrueFraction is the accurate way to convert floats. Looking good would mean rounded nicely.
Maybe I'm expecting printShowingDecimalPlaces: to be something it isn't intended to be but I don't see any other similar method that would perform the pretty way.
Lou
On Tue, 26 Mar 2013 22:11:41 +0100, Nicolas Cellier nicolas.cellier.aka.nice@gmail.com wrote:
2013/3/26 Louis LaBrunda Lou@keystone-software.com:
Hi Nicolas,
snip...
In VA Smalltalk:
0.995 = 0.995s3 => true 0.995 < 0.995s3 => false
This is a default st80 behavior which converts minimumGenerality number to maximumGenerality.
In VA when one converts 0.995 to decimal, one gets 0.995s15, which when compared to 0.995s3 is equal. And when 0.995s3 is converted to float that float compares equal to 0.995.
Maybe some of these conventions are hardwired in VM, I don't know, but in other dialects it's handled at image side.
The idea was this one: if you perform an operation between an inexact value and an exact value, the result is inexact. So Float + * / - Integer,Fraction,ScaledDecimal will result into a Float. Thus Float got a higher generality.
This would imply that one should convert the ScaledDecimal to a Float before the compare. But as above that would not change things (I think maybe because VA uses 8 byte floats).
This is from Squeak:
0.995s3 asFloat 0.995
0.995s3 asFloat = 0.995 true
0.995 asScaledDecimal 0.99499999s8
(995 / 1000) asScaledDecimal 0.995s3
0.995 asTrueFraction asScaledDecimal 0.99499999999999999555910790149937383830547332763671875s53
(0.995 * 1000000000) asScaledDecimal / 1000000000 0.99500000s8
Yep, but (0.995 * 1000000000) is inexact... You can check that (0.995 * 1000000000) ~~ (0.995 asFraction * 1000000000)
You cannot rely on results of any such inexact calculus, or you'll be building on sand.
0.995 * 1000000000 9.95e8
Sorry, but I have run out of time to play at the moment. So, I will just throw a thought out there. I think there may be a problem with asTrueFraction. Which if implemented differently might not make 0.995 < 0.995s3.
Well I doubt, but I'm all ears ;)
(0.995 asFraction storeStringBase: 2) -> '(2r11111110101110000101000111101011100001010001111010111/2r100000000000000000000000000000000000000000000000000000)'.
(0.995 successor) asFraction storeStringBase: 2 -> '(2r11111110101110000101000111101011100001010001111011/2r100000000000000000000000000000000000000000000000000)' -> '(2r11111110101110000101000111101011100001010001111011000/2r100000000000000000000000000000000000000000000000000000)'.
(0.995 predecessor) asFraction storeStringBase: 2 -> '(2r1111111010111000010100011110101110000101000111101011/2r10000000000000000000000000000000000000000000000000000)' -> '(2r11111110101110000101000111101011100001010001111010110/2r100000000000000000000000000000000000000000000000000000)'.
0.995 ulp asFraction storeStringBase: 2 -> '(2r1/2r100000000000000000000000000000000000000000000000000000)'.
(995/1000 - 0.995 asFraction) / 0.995 ulp -> 0.04.
(995/1000 - 0.995 successor asFraction) / 0.995 ulp -> -0.96.
(995/1000 - 0.995 successor asFraction) / 0.995 ulp -> 1.04.
and numerator/denominator have this property:
2r11111110101110000101000111101011100001010001111010111 highBit -> 53. 2r100000000000000000000000000000000000000000000000000000 highBit -> 54.
So, 53 bits of significand, a power of two on denominator, that sounds like a correct Float, slightly smaller than 1. Next significand and previous significand are both further of (995/1000) than 0.995 is. 0.995 is closest Float to 995/1000 and is smaller than 995/1000, no doubt.
Nicolas
The idea behind Squeak change is that every Float has an exact value (asTrueFraction). Since we have only two possible answers true/false for comparison, and no maybe or dontKnow, it makes sense to compare the exact value. This reduces the number of paradoxal equalities
| a b c | a := 1 << Float precision. b := a + 1. c := a asFloat. { c = b. c = a. a = b. }
In VW and VA, the first two are true, the last is false, which suggests that = is not an equivalence relation. In Squeak, only the second one is true.
Same with inequalities, we expect (a < b) & (a = c) ==> (c < b) etc... This is still true in Squeak, not in VA/VW/st80. I think gst and Dolphin and maybe stx adopted Squeak behavior (to be confirmed).
I think in VA Smalltalk there is some VM magic going on that makes the above work the way it does. The #= and #< of Float are implemented in a primitive. I guess when converting the '0.995' string to a float a little is lost and that would make it less than 0.995s3 but there is a lot of code floating around (sorry about the puns) to make floats look better. In that case I would think people would want (0.995 printShowingDecimalPlaces: 2) to show '1.00' and not '0.99'.
No, people should not rely on such behavior with Floats, because sooner than later they will be bitten. We cannot cheat very long with this kind of assumptions. My POV is that it is better to educate about false expectations with Float, and that's the main benefit of (1/10) ~= 0.1.
I would add that such print policy is not the behaviour of every other language I know of.
printf( "%.2f",0.995) -> 0.99
Because libm are carefully written nowdays and other languages libraries are either built over libm or much more careful than Smalltalk were (that mostly means more recent).
Anyway, we should try not to use floats without a very, very good reason. Like we have to send them outside of Smalltalk or we really need the speed or decimals and fractions take up too much memory. But then we must live with their inaccuracies and display mess.
Good advice, let's put those expectations on decimal fractions (ScaledDecimal/FixedPoint) or general Fraction.
I have an untested theory that fractions can be close in speed to floats because divisions (that are expensive) can be pushed to the end of a computation because with fractions they are multiplies.
Lou
Why not, but huge numerators and denominators are not cheap compared to Float operations. OK reducing the fraction is the expensive part, but how does the cost grow versus the length of operands?
Also some geometric operations are not even possible on Q (hypot) so you might soon need AlgebraicNumber. And Smalltalk is also about graphics and geometry.
Nicolas
Louis LaBrunda Keystone Software Corp. SkypeMe callto://PhotonDemon mailto:Lou@Keystone-Software.com http://www.Keystone-Software.com
Louis LaBrunda Keystone Software Corp. SkypeMe callto://PhotonDemon mailto:Lou@Keystone-Software.com http://www.Keystone-Software.com
----------------------------------------------------------- Louis LaBrunda Keystone Software Corp. SkypeMe callto://PhotonDemon mailto:Lou@Keystone-Software.com http://www.Keystone-Software.com
Well, that's what I'm fighting against: looking good. The good looking float are putting plenty of traps on our path, and it's better to know about it. I say that it's impossible to maintain the illusion of good looking and friendly float, and in that case it's better to tell the awfull truth. The changes in Squeak/Pharo are here to un-learn what we thought about decimal numbers, or more accurately to learn that these rules don't apply exactly to Float.
Beside, in most languages, if you round the float/double 0.995 to 2 digits - either by a printf or another operation like round(0.995 , 2) - you'll get 0.99 not 1.00.
Nicolas
2013/3/29 Louis LaBrunda Lou@keystone-software.com
Hi Nicolas,
Sorry to be getting back to this so late. I take back what I said about asTrueFraction. I'm sure it is fine. But I think I question its use in this case.
This all started when Stéphane said this:
0.995 printShowingDecimalPlaces: 2 ==> '1.00'
was wrong and I didn't think so.
When one uses a method like printShowingDecimalPlaces: one expects (maybe there is some other similar method that one should expect this) the result to look good (yes that is very subjective) but I would say looking good is more important than being accurate. Yes, that is an admission that asTrueFraction is the accurate way to convert floats. Looking good would mean rounded nicely.
Maybe I'm expecting printShowingDecimalPlaces: to be something it isn't intended to be but I don't see any other similar method that would perform the pretty way.
Lou
On Tue, 26 Mar 2013 22:11:41 +0100, Nicolas Cellier nicolas.cellier.aka.nice@gmail.com wrote:
2013/3/26 Louis LaBrunda Lou@keystone-software.com:
Hi Nicolas,
snip...
In VA Smalltalk:
0.995 = 0.995s3 => true 0.995 < 0.995s3 => false
This is a default st80 behavior which converts minimumGenerality number to maximumGenerality.
In VA when one converts 0.995 to decimal, one gets 0.995s15, which when compared to 0.995s3 is equal. And when 0.995s3 is converted to float
that
float compares equal to 0.995.
Maybe some of these conventions are hardwired in VM, I don't know, but in other dialects it's handled at image side.
The idea was this one: if you perform an operation between an inexact value and an exact value, the result is inexact. So Float + * / - Integer,Fraction,ScaledDecimal will result into a
Float.
Thus Float got a higher generality.
This would imply that one should convert the ScaledDecimal to a Float before the compare. But as above that would not change things (I think maybe because VA uses 8 byte floats).
This is from Squeak:
0.995s3 asFloat 0.995
0.995s3 asFloat = 0.995 true
0.995 asScaledDecimal 0.99499999s8
(995 / 1000) asScaledDecimal 0.995s3
0.995 asTrueFraction asScaledDecimal 0.99499999999999999555910790149937383830547332763671875s53
(0.995 * 1000000000) asScaledDecimal / 1000000000 0.99500000s8
Yep, but (0.995 * 1000000000) is inexact... You can check that (0.995 * 1000000000) ~~ (0.995 asFraction * 1000000000)
You cannot rely on results of any such inexact calculus, or you'll be building on sand.
0.995 * 1000000000 9.95e8
Sorry, but I have run out of time to play at the moment. So, I will
just
throw a thought out there. I think there may be a problem with asTrueFraction. Which if implemented differently might not make 0.995 < 0.995s3.
Well I doubt, but I'm all ears ;)
(0.995 asFraction storeStringBase: 2) ->
'(2r11111110101110000101000111101011100001010001111010111/2r100000000000000000000000000000000000000000000000000000)'.
(0.995 successor) asFraction storeStringBase: 2 ->
'(2r11111110101110000101000111101011100001010001111011/2r100000000000000000000000000000000000000000000000000)'
->
'(2r11111110101110000101000111101011100001010001111011000/2r100000000000000000000000000000000000000000000000000000)'.
(0.995 predecessor) asFraction storeStringBase: 2 ->
'(2r1111111010111000010100011110101110000101000111101011/2r10000000000000000000000000000000000000000000000000000)'
->
'(2r11111110101110000101000111101011100001010001111010110/2r100000000000000000000000000000000000000000000000000000)'.
0.995 ulp asFraction storeStringBase: 2 -> '(2r1/2r100000000000000000000000000000000000000000000000000000)'.
(995/1000 - 0.995 asFraction) / 0.995 ulp -> 0.04.
(995/1000 - 0.995 successor asFraction) / 0.995 ulp -> -0.96.
(995/1000 - 0.995 successor asFraction) / 0.995 ulp -> 1.04.
and numerator/denominator have this property:
2r11111110101110000101000111101011100001010001111010111 highBit -> 53. 2r100000000000000000000000000000000000000000000000000000 highBit -> 54.
So, 53 bits of significand, a power of two on denominator, that sounds like a correct Float, slightly smaller than 1. Next significand and previous significand are both further of (995/1000) than 0.995 is. 0.995 is closest Float to 995/1000 and is smaller than 995/1000, no doubt.
Nicolas
The idea behind Squeak change is that every Float has an exact value (asTrueFraction). Since we have only two possible answers true/false for comparison, and no maybe or dontKnow, it makes sense to compare the exact value. This reduces the number of paradoxal equalities
| a b c | a := 1 << Float precision. b := a + 1. c := a asFloat. { c = b. c = a. a = b. }
In VW and VA, the first two are true, the last is false, which suggests that = is not an equivalence relation. In Squeak, only the second one is true.
Same with inequalities, we expect (a < b) & (a = c) ==> (c < b) etc... This is still true in Squeak, not in VA/VW/st80. I think gst and Dolphin and maybe stx adopted Squeak behavior (to be
confirmed).
I think in VA Smalltalk there is some VM magic going on that makes the above work the way it does. The #= and #< of Float are implemented
in a
primitive. I guess when converting the '0.995' string to a float a
little
is lost and that would make it less than 0.995s3 but there is a lot
of code
floating around (sorry about the puns) to make floats look better.
In that
case I would think people would want (0.995
printShowingDecimalPlaces: 2)
to show '1.00' and not '0.99'.
No, people should not rely on such behavior with Floats, because sooner than later they will be bitten. We cannot cheat very long with this kind of assumptions. My POV is that it is better to educate about false expectations with Float, and that's the main benefit of (1/10) ~= 0.1.
I would add that such print policy is not the behaviour of every other language I know of.
printf( "%.2f",0.995) -> 0.99
Because libm are carefully written nowdays and other languages libraries are either built over libm or much more careful than Smalltalk were (that mostly means more recent).
Anyway, we should try not to use floats without a very, very good
reason.
Like we have to send them outside of Smalltalk or we really need the
speed
or decimals and fractions take up too much memory. But then we must
live
with their inaccuracies and display mess.
Good advice, let's put those expectations on decimal fractions (ScaledDecimal/FixedPoint) or general Fraction.
I have an untested theory that fractions can be close in speed to
floats
because divisions (that are expensive) can be pushed to the end of a computation because with fractions they are multiplies.
Lou
Why not, but huge numerators and denominators are not cheap compared to Float operations. OK reducing the fraction is the expensive part, but how does the cost grow versus the length of operands?
Also some geometric operations are not even possible on Q (hypot) so you might soon need AlgebraicNumber. And Smalltalk is also about graphics and geometry.
Nicolas
Louis LaBrunda Keystone Software Corp. SkypeMe callto://PhotonDemon mailto:Lou@Keystone-Software.com http://www.Keystone-Software.com
Louis LaBrunda Keystone Software Corp. SkypeMe callto://PhotonDemon mailto:Lou@Keystone-Software.com http://www.Keystone-Software.com
Louis LaBrunda Keystone Software Corp. SkypeMe callto://PhotonDemon mailto:Lou@Keystone-Software.com http://www.Keystone-Software.com
It also is good for remembering presentation must be able to be distinct from internal representation.
On Fri, Mar 29, 2013 at 1:41 PM, Nicolas Cellier nicolas.cellier.aka.nice@gmail.com wrote:
Well, that's what I'm fighting against: looking good. The good looking float are putting plenty of traps on our path, and it's better to know about it. I say that it's impossible to maintain the illusion of good looking and friendly float, and in that case it's better to tell the awfull truth. The changes in Squeak/Pharo are here to un-learn what we thought about decimal numbers, or more accurately to learn that these rules don't apply exactly to Float.
Beside, in most languages, if you round the float/double 0.995 to 2 digits - either by a printf or another operation like round(0.995 , 2) - you'll get 0.99 not 1.00.
Nicolas
2013/3/29 Louis LaBrunda Lou@keystone-software.com
Hi Nicolas,
Sorry to be getting back to this so late. I take back what I said about asTrueFraction. I'm sure it is fine. But I think I question its use in this case.
This all started when Stéphane said this:
0.995 printShowingDecimalPlaces: 2 ==> '1.00'
was wrong and I didn't think so.
When one uses a method like printShowingDecimalPlaces: one expects (maybe there is some other similar method that one should expect this) the result to look good (yes that is very subjective) but I would say looking good is more important than being accurate. Yes, that is an admission that asTrueFraction is the accurate way to convert floats. Looking good would mean rounded nicely.
Maybe I'm expecting printShowingDecimalPlaces: to be something it isn't intended to be but I don't see any other similar method that would perform the pretty way.
Lou
On Tue, 26 Mar 2013 22:11:41 +0100, Nicolas Cellier nicolas.cellier.aka.nice@gmail.com wrote:
2013/3/26 Louis LaBrunda Lou@keystone-software.com:
Hi Nicolas,
snip...
In VA Smalltalk:
0.995 = 0.995s3 => true 0.995 < 0.995s3 => false
This is a default st80 behavior which converts minimumGenerality number to maximumGenerality.
In VA when one converts 0.995 to decimal, one gets 0.995s15, which when compared to 0.995s3 is equal. And when 0.995s3 is converted to float that float compares equal to 0.995.
Maybe some of these conventions are hardwired in VM, I don't know, but in other dialects it's handled at image side.
The idea was this one: if you perform an operation between an inexact value and an exact value, the result is inexact. So Float + * / - Integer,Fraction,ScaledDecimal will result into a Float. Thus Float got a higher generality.
This would imply that one should convert the ScaledDecimal to a Float before the compare. But as above that would not change things (I think maybe because VA uses 8 byte floats).
This is from Squeak:
0.995s3 asFloat 0.995
0.995s3 asFloat = 0.995 true
0.995 asScaledDecimal 0.99499999s8
(995 / 1000) asScaledDecimal 0.995s3
0.995 asTrueFraction asScaledDecimal 0.99499999999999999555910790149937383830547332763671875s53
(0.995 * 1000000000) asScaledDecimal / 1000000000 0.99500000s8
Yep, but (0.995 * 1000000000) is inexact... You can check that (0.995 * 1000000000) ~~ (0.995 asFraction * 1000000000)
You cannot rely on results of any such inexact calculus, or you'll be building on sand.
0.995 * 1000000000 9.95e8
Sorry, but I have run out of time to play at the moment. So, I will just throw a thought out there. I think there may be a problem with asTrueFraction. Which if implemented differently might not make 0.995 < 0.995s3.
Well I doubt, but I'm all ears ;)
(0.995 asFraction storeStringBase: 2) -> '(2r11111110101110000101000111101011100001010001111010111/2r100000000000000000000000000000000000000000000000000000)'.
(0.995 successor) asFraction storeStringBase: 2 -> '(2r11111110101110000101000111101011100001010001111011/2r100000000000000000000000000000000000000000000000000)' -> '(2r11111110101110000101000111101011100001010001111011000/2r100000000000000000000000000000000000000000000000000000)'.
(0.995 predecessor) asFraction storeStringBase: 2 -> '(2r1111111010111000010100011110101110000101000111101011/2r10000000000000000000000000000000000000000000000000000)' -> '(2r11111110101110000101000111101011100001010001111010110/2r100000000000000000000000000000000000000000000000000000)'.
0.995 ulp asFraction storeStringBase: 2 -> '(2r1/2r100000000000000000000000000000000000000000000000000000)'.
(995/1000 - 0.995 asFraction) / 0.995 ulp -> 0.04.
(995/1000 - 0.995 successor asFraction) / 0.995 ulp -> -0.96.
(995/1000 - 0.995 successor asFraction) / 0.995 ulp -> 1.04.
and numerator/denominator have this property:
2r11111110101110000101000111101011100001010001111010111 highBit -> 53. 2r100000000000000000000000000000000000000000000000000000 highBit -> 54.
So, 53 bits of significand, a power of two on denominator, that sounds like a correct Float, slightly smaller than 1. Next significand and previous significand are both further of (995/1000) than 0.995 is. 0.995 is closest Float to 995/1000 and is smaller than 995/1000, no doubt.
Nicolas
The idea behind Squeak change is that every Float has an exact value (asTrueFraction). Since we have only two possible answers true/false for comparison, and no maybe or dontKnow, it makes sense to compare the exact value. This reduces the number of paradoxal equalities
| a b c | a := 1 << Float precision. b := a + 1. c := a asFloat. { c = b. c = a. a = b. }
In VW and VA, the first two are true, the last is false, which suggests that = is not an equivalence relation. In Squeak, only the second one is true.
Same with inequalities, we expect (a < b) & (a = c) ==> (c < b) etc... This is still true in Squeak, not in VA/VW/st80. I think gst and Dolphin and maybe stx adopted Squeak behavior (to be confirmed).
I think in VA Smalltalk there is some VM magic going on that makes the above work the way it does. The #= and #< of Float are implemented in a primitive. I guess when converting the '0.995' string to a float a little is lost and that would make it less than 0.995s3 but there is a lot of code floating around (sorry about the puns) to make floats look better. In that case I would think people would want (0.995 printShowingDecimalPlaces: 2) to show '1.00' and not '0.99'.
No, people should not rely on such behavior with Floats, because sooner than later they will be bitten. We cannot cheat very long with this kind of assumptions. My POV is that it is better to educate about false expectations with Float, and that's the main benefit of (1/10) ~= 0.1.
I would add that such print policy is not the behaviour of every other language I know of.
printf( "%.2f",0.995) -> 0.99
Because libm are carefully written nowdays and other languages libraries are either built over libm or much more careful than Smalltalk were (that mostly means more recent).
Anyway, we should try not to use floats without a very, very good reason. Like we have to send them outside of Smalltalk or we really need the speed or decimals and fractions take up too much memory. But then we must live with their inaccuracies and display mess.
Good advice, let's put those expectations on decimal fractions (ScaledDecimal/FixedPoint) or general Fraction.
I have an untested theory that fractions can be close in speed to floats because divisions (that are expensive) can be pushed to the end of a computation because with fractions they are multiplies.
Lou
Why not, but huge numerators and denominators are not cheap compared to Float operations. OK reducing the fraction is the expensive part, but how does the cost grow versus the length of operands?
Also some geometric operations are not even possible on Q (hypot) so you might soon need AlgebraicNumber. And Smalltalk is also about graphics and geometry.
Nicolas
Louis LaBrunda Keystone Software Corp. SkypeMe callto://PhotonDemon mailto:Lou@Keystone-Software.com http://www.Keystone-Software.com
Louis LaBrunda Keystone Software Corp. SkypeMe callto://PhotonDemon mailto:Lou@Keystone-Software.com http://www.Keystone-Software.com
Louis LaBrunda Keystone Software Corp. SkypeMe callto://PhotonDemon mailto:Lou@Keystone-Software.com http://www.Keystone-Software.com
2013/3/26 Louis LaBrunda Lou@keystone-software.com:
Hi Nicolas,
snip...
In VA Smalltalk:
0.995 = 0.995s3 => true 0.995 < 0.995s3 => false
This is a default st80 behavior which converts minimumGenerality number to maximumGenerality.
In VA when one converts 0.995 to decimal, one gets 0.995s15, which when compared to 0.995s3 is equal. And when 0.995s3 is converted to float that float compares equal to 0.995.
Maybe some of these conventions are hardwired in VM, I don't know, but in other dialects it's handled at image side.
The idea was this one: if you perform an operation between an inexact value and an exact value, the result is inexact. So Float + * / - Integer,Fraction,ScaledDecimal will result into a Float. Thus Float got a higher generality.
This would imply that one should convert the ScaledDecimal to a Float before the compare. But as above that would not change things (I think maybe because VA uses 8 byte floats).
No, that's precisely what st80 VW and VA did, convert exact->inexact (asFloat)
And what Squeak/gst/Dolphin/stx don't: It's better to convert Inexact->exact and perform the comparison exactly for preserving some of = and < properties.
Having inexact numbers is not a license to spoil accuracy. If you look at IEEE754 standard, results are always equivalent to an exact operation followed by a correct rounding (to nearest float, tie to even in default mode). Even if they are inexact, we perform the operations as if they were exact (that's why Float differs from interval calculus).
OK, the exception is transcendental functions which can be very expensive to compute accurately, so the standard have relaxed the error to 3 ulp or something like that... But * + - / sqrt are accurate to 1/2 ulp.
Nicolas
This is from Squeak:
0.995s3 asFloat 0.995
0.995s3 asFloat = 0.995 true
0.995 asScaledDecimal 0.99499999s8
(995 / 1000) asScaledDecimal 0.995s3
0.995 asTrueFraction asScaledDecimal 0.99499999999999999555910790149937383830547332763671875s53
(0.995 * 1000000000) asScaledDecimal / 1000000000 0.99500000s8
0.995 * 1000000000 9.95e8
Sorry, but I have run out of time to play at the moment. So, I will just throw a thought out there. I think there may be a problem with asTrueFraction. Which if implemented differently might not make 0.995 < 0.995s3.
The idea behind Squeak change is that every Float has an exact value (asTrueFraction). Since we have only two possible answers true/false for comparison, and no maybe or dontKnow, it makes sense to compare the exact value. This reduces the number of paradoxal equalities
| a b c | a := 1 << Float precision. b := a + 1. c := a asFloat. { c = b. c = a. a = b. }
In VW and VA, the first two are true, the last is false, which suggests that = is not an equivalence relation. In Squeak, only the second one is true.
Same with inequalities, we expect (a < b) & (a = c) ==> (c < b) etc... This is still true in Squeak, not in VA/VW/st80. I think gst and Dolphin and maybe stx adopted Squeak behavior (to be confirmed).
I think in VA Smalltalk there is some VM magic going on that makes the above work the way it does. The #= and #< of Float are implemented in a primitive. I guess when converting the '0.995' string to a float a little is lost and that would make it less than 0.995s3 but there is a lot of code floating around (sorry about the puns) to make floats look better. In that case I would think people would want (0.995 printShowingDecimalPlaces: 2) to show '1.00' and not '0.99'.
No, people should not rely on such behavior with Floats, because sooner than later they will be bitten. We cannot cheat very long with this kind of assumptions. My POV is that it is better to educate about false expectations with Float, and that's the main benefit of (1/10) ~= 0.1.
I would add that such print policy is not the behaviour of every other language I know of.
printf( "%.2f",0.995) -> 0.99
Because libm are carefully written nowdays and other languages libraries are either built over libm or much more careful than Smalltalk were (that mostly means more recent).
Anyway, we should try not to use floats without a very, very good reason. Like we have to send them outside of Smalltalk or we really need the speed or decimals and fractions take up too much memory. But then we must live with their inaccuracies and display mess.
Good advice, let's put those expectations on decimal fractions (ScaledDecimal/FixedPoint) or general Fraction.
I have an untested theory that fractions can be close in speed to floats because divisions (that are expensive) can be pushed to the end of a computation because with fractions they are multiplies.
Lou
Why not, but huge numerators and denominators are not cheap compared to Float operations. OK reducing the fraction is the expensive part, but how does the cost grow versus the length of operands?
Also some geometric operations are not even possible on Q (hypot) so you might soon need AlgebraicNumber. And Smalltalk is also about graphics and geometry.
Nicolas
Louis LaBrunda Keystone Software Corp. SkypeMe callto://PhotonDemon mailto:Lou@Keystone-Software.com http://www.Keystone-Software.com
Louis LaBrunda Keystone Software Corp. SkypeMe callto://PhotonDemon mailto:Lou@Keystone-Software.com http://www.Keystone-Software.com
squeak-dev@lists.squeakfoundation.org