Difference between revisions of "User:Regina/MyDrafts"
(sensitivity of trigonometric functions) |
((checkpoint save)) |
||
| Line 1: | Line 1: | ||
''' Accuracy ''' | ''' Accuracy ''' | ||
| − | This article is about situations, where you think, that Calc calculates wrong. You might have found a bug, and then you should write an issue. But look here first; perhaps the unexpected results | + | This article is about situations, where you think, that Calc calculates wrong. You might have found a bug, and then you should write an issue. But look here first; perhaps the unexpected results have reasons, you haven't been aware. |
== Precision in Calc == | == Precision in Calc == | ||
| − | Calc uses for its calculation floating point numbers in double precision as defined in IEEE 754 standard. You get the | + | Calc uses for its calculation floating point numbers in double precision as defined in IEEE 754 standard. You get the full representation in a spreadsheet cell using the scientific format with format code <tt>0.00000000000000E+000</tt>. |
Calc rounds to two decimals in the default settings. | Calc rounds to two decimals in the default settings. | ||
| − | + | ToDo: Explain setting "precision as shown". | |
| − | + | ||
| − | + | ||
| − | + | ||
But although you can force Calc to show 15 decimal digits, these might not be all accurate. The following sections list some of the problems. | But although you can force Calc to show 15 decimal digits, these might not be all accurate. The following sections list some of the problems. | ||
| Line 18: | Line 15: | ||
== Converting between Number Systems == | == Converting between Number Systems == | ||
| − | + | Most non integer numbers have infinite decimal places in binary format, which have to be rounded somewhere. | |
| − | + | For example, the number 0.1 cannot be written exactly in binary format. Because of this rounding it can happen, that two mathematical different numbers have the same (rounded) internal binary value. | |
| − | Most non integer numbers have infinite decimal places in binary format, which have to be rounded somewhere. | + | The other way round, the decimal values are rounded too. So you will see no difference between =SQRT(2) and =SQRT(2)+9E-15, but EXP(SQRT(2)) and EXP(SQRT(2)+9E-15) differs in the last shown figure. |
| − | + | ||
| − | + | ||
| − | + | ||
| − | + | ||
| − | + | ||
| − | + | ||
| − | + | ||
| − | + | ||
== No Symbolic π == | == No Symbolic π == | ||
Revision as of 15:17, 20 May 2010
Accuracy
This article is about situations, where you think, that Calc calculates wrong. You might have found a bug, and then you should write an issue. But look here first; perhaps the unexpected results have reasons, you haven't been aware.
Contents
Precision in Calc
Calc uses for its calculation floating point numbers in double precision as defined in IEEE 754 standard. You get the full representation in a spreadsheet cell using the scientific format with format code 0.00000000000000E+000.
Calc rounds to two decimals in the default settings.
ToDo: Explain setting "precision as shown".
But although you can force Calc to show 15 decimal digits, these might not be all accurate. The following sections list some of the problems.
Converting between Number Systems
Most non integer numbers have infinite decimal places in binary format, which have to be rounded somewhere. For example, the number 0.1 cannot be written exactly in binary format. Because of this rounding it can happen, that two mathematical different numbers have the same (rounded) internal binary value. The other way round, the decimal values are rounded too. So you will see no difference between =SQRT(2) and =SQRT(2)+9E-15, but EXP(SQRT(2)) and EXP(SQRT(2)+9E-15) differs in the last shown figure.
No Symbolic π
From mathematics you know sin(π)= 0 and you know that tan(π/2) is undefined. But you cannot get these results in Calc, because the value π is always treated as rounded floating point number. It makes no difference using PI() or RADIANDS(180). Calc cannot evaluate π symbolically as computer algebra systems do. That is no special limitation of Calc, but other often used spreadsheet applications work only numerically too.
| A | B | C | |
|---|---|---|---|
| 1 | 1.63317787283838E+016 | =TAN(PI()/2) | |
| 2 | 1.22460635382238E-016 | =SIN(RADIANS(180)) | |
| 3 |
No Fractional Arithmetic
Calc kann zwar Zahlen als Brüche darstellen, aber nicht mit ihnen rechnen.
Cancellation
If you subtract two non integer numbers, which have nearly the same value, the result has less significant digits then the initial values.
| A | B | C | |
|---|---|---|---|
| 1 | 9.99411764795882E-001 | =0.999411764795882 | |
| 2 | 9.99411764705882E-001 | =1699/1700 | |
| 3 | 8.99997854020285E-011 | =A1-A2 | |
| 4 | 8.99996470588235E-011 |
Cell A4 shows the correct result of 
, calculated with a computer algebra system with high precision.
| A | B | C | |
|---|---|---|---|
| 1 | 0.99999876543210000000 | ||
| 2 | 0.00000123456790002141 | =1-A1 | |
| 3 | |||
| 4 |
Calculating manually gives 1−0.99999876543210000000 = 0.00000123456790000000
Ill-conditioned problems
Verhalten bei Polstellen
Sensitivity
Stabilität des Algorithmus Todo
Hidden Bits
How much changes the result, if the input varies with one, two, three... bit in the internal representation?
Trigonometric functions are very sensitive for huge input values.
Example of formula imput
| A | B | C | |
|---|---|---|---|
| 1 | =2^44-0.004 | =sin(A1) | |
| 2 | =2^44 | =sin(A2) | |
| 3 | =2^44+0.004 | =sin(A3) | |
| 4 |
Results shown in 15 figures precision
| A | B | C | |
|---|---|---|---|
| 1 | 17592186044416.00 | 0.386569623644289 | |
| 2 | 17592186044416.00 | 0.390169244205272 | |
| 3 | 17592186044416.00 | 0.393762911263629 | |
| 4 |
The correct value rounded to 15 figures precision would be sin(2^44) = 0.390169223351877
The trigonometric functions are reliable for arguments up to 2^27 with at least 13 figures accuracy. For larger values the accuracy decreases to about 6 figures accuracy for integral arguments and 2 figures accuracy for fractional values reaching 2^44. Huge arguments in the area above 2^44 might be totally wrong. If you need trigonometric functions for huge arguments, a spreadsheet is the wrong application for you.
The trigonometric functions are similar sensitive for arguments near their roots, but for those arguments the absolute result is near zero. So you will not notice the error, if you work with rounding to 15 or less decimal places.
