Related
static_casting from a floating point to an integer simply strips the fractional point of the number. For example static_cast<int>(13.9999999) yields 13.
Not all integers are representable as floating point numbers. For example internally the closest float to 13,000,000 may be: 12999999.999999.
In this hypothetical case, I'd expect to get an unexpected result from:
const auto foo = 12'999'999.5F;
const auto bar = static_cast<long long>(ceil(foo));
My assumption is that such a breakdown does occur at some point, if not necessarily at 13,000,000. I'd just like to know the range over which I can trust static_cast<long long>(ceif(foo))?
For example internally the closest float to 13,000,000 may be: 12999999.999999.
That is not possible in any normal floating-point format. The floating-point representation of numbers is equivalent to M•be, where b is a fixed base (e.g., 2 for binary floating-point) and M and e are integers with some restrictions on their values. In order for a value like 13,000,000-x to be represented, where x is some positive value less than 1, e must be negative (because M•be for a non-negative e is an integer). If so, then M•b0 is an integer larger than M•be, so it is larger than 13,000,000, and so 13,000,000 can be represented as M'•b0, where M' is a positive integer less than M and hence fits in the range of allowed values for M (in any normal floating-point format). (Perhaps some bizarre floating-point format might impose a strange range on M or e that prevents this, but no normal format does.)
Regarding your code:
auto test = 0LL;
const auto floater = 0.5F;
for(auto i = 0LL; i == test; i = std::ceil(i + floater)) ++test;
cout << test << endl;
When i was 8,388,608, the mathematical result of 8,388,608 + .5 is 8,388,608.5. This is not representable in the float format on your system, so it was rounded to 8,388,608. The ceil of this is 8,388,608. At this point, test was 8,388,609, so the loop stopped. So this code does not demonstrate that 8,388,608.5 is representable and 8,388,609 is not.
Behavior seems to return to normal if I do: ceil(8'388'609.5F) which will correctly return 8,388,610.
8,388,609.5 is not representable in the float format on your system, so it was rounded by the rule “round to nearest, ties to even.” The two nearest representable values are 8,388,609, and 8,388,610. Since they are equally far apart, the result was 8,388,610. That value was passed to ceil, which of course returned 8,388,610.
On Visual Studio 2015 I got 8,388,609 which is a horrifying small safe range.
In the IEEE-754 basic 32-bit binary format, all integers from -16,777,216 to +16,777,216 are representable, because the format has a 24-bit significand.
Floating point numbers are represented by 3 integers, cbq where:
c is the mantissa (so for the number: 12,999,999.999999 c would be 12,999,999,999,999)
q is the exponent (so for the number: 12,999,999.999999 q would be -6)
b is the base (IEEE-754 requires b to be either 10 or 2; in the representation above b is 10)
From this it's easy to see that a floating point with the capability of representing 12,999,999.999999 also has the capability of representing 13,000,000.000000 using a c of 1,300,000,000,000 and a q of -5.
This example is a bit contrived in that the chosen b is 10, where in almost all implementations the chosen base is 2. But it's worth pointing out that even with a b of 2 the q functions as a shift left or right of the mantissa.
Next let's talk about a range here. Obviously a 32-bit floating point cannot represent all the integers represented by a 32-bit integer, as the floating point must also represent so many much larger or smaller numbers. Since the exponent is simply shifting the mantissa, a floating point number can always exactly represent every integer that can be represented by it's mantissa. Given the traditional IEEE-754 binary base floating point numbers:
A 32-bit (float) has a 24-bit mantissa so it can represent all integers in the range [-16,777,215, 16,777,215]
A 64-bit (double) has a 53-bit mantissa so it can represent all integers in the range [-9,007,199,254,740,991, 9,007,199,254,740,991]
A 128-bit (long double depending upon implementation) has a 113-bit mantissa so it can represent all integers in the range [-103,845,937,170,696,552,570,609,926,584,40,191, 103,845,937,170,696,552,570,609,926,584,40,191]
[source]
c++ provides digits as a method of finding this number for a given floating point type. (Though admittedly even a long long is too small to represent a 113-bit mantissa.) For example a float's maximum mantissa could be found by:
(1LL << numeric_limits<float>::digits) - 1LL
Having thoroughly explained the mantissa, let's revisit the exponent section to talk about how a floating point is actually stored. Take 13,000,000.0 that could be represented as:
c = 13, q = 6, b = 10
c = 130, q = 5, b = 10
c = 1,300, q = 4, b = 10
And so on. For the traditional binary format IEEE-754 requires:
The representation is made unique by choosing the smallest representable exponent that retains the most significant bit (MSB) within the selected word size and format. Further, the exponent is not represented directly, but a bias is added so that the smallest representable exponent is represented as 1, with 0 used for subnormal numbers
To explain this in the more familiar base-10 if our mantissa has 14 decimal places, the implementation would look like this:
c = 13,000,000,000,000 so the MSB will be used in the represented number
q = 6 This is a little confusing, it's cause of the bias introduced here; logically q = -6 but the bias is set so that when q = 0 only the MSB of c is immediately to the left of the decimal point, meaning that c = 13,000,000,000,000, q = 0, b = 10 will represent 1.3
b = 10 again the above rules are really only required for base-2 but I've shown them as they would apply to base-10 for the purpose of explaination
Translated back to base-2 this means that a q of numeric_limits<T>::digits - 1 has only zeros after the decimal place. ceil only has an effect if there is a fractional part of the number.
A final point of explanation here, is the range over which ceil will have an effect. After the exponent of a floating point is larger than numeric_limits<T>::digits continuing to increase it only introduces trailing zeros to the resulting number, thus calling ceil when q is greater than or equal to numeric_limits<T>::digits - 2LL. And since we know the MSB of c will be used in the number this means that c must be smaller than (1LL << numeric_limits<T>::digits - 1LL) - 1LL Thus for ceil to have an effect on the traditional binary IEEE-754 floating point:
A 32-bit (float) must be smaller than 8,388,607
A 64-bit (double) must be smaller than 4,503,599,627,370,495
A 128-bit (long double depending upon implementation) must be smaller than 5,192,296,858,534,827,628,530,496,329,220,095
This question already has answers here:
What range of numbers can be represented in a 16-, 32- and 64-bit IEEE-754 systems?
(7 answers)
Closed 8 years ago.
I'm trying to make a function that enables me to print floats.
Right now, I'm encountering two strange behaviors :
Sometimes, values like 1.3 come out as 1.2999999 instead of 1.3000000,and sometimes values like 1.234567 come out as 1.2345672 instead of 1.2345670.
Here's the source code :
int ft_putflt(float f)
{
int ret;
int intpart;
int i;
ret = 0;
i = 0;
intpart = (int)f;
ft_putnbr(intpart);
ret = ft_nbrlen(intpart) + 8;
write(1, ".", 1);
while (i++ < 7)
{
f *= 10;
ft_putchar(48 + ((int)f % 10));
}
return (ret);
}
ft_putnbr is OK AFAIK.
ft_putchar is a simple call to "write(1, &c, 1)".
test values (value : output)
1.234567 : 1.2345672 (!)
1.2345670 : 1.2345672 (!)
1.0000001 : 1.0000001 OK
0.1234567 : 0.1234567 OK
0.67 : 0.6700000 OK
1.3 : 1.3000000 OK (fixed it)
1.321012 : 1.3210119 (!)
1.3210121 : 1.3210122 (!)
This all seems a bit mystic to me... Loss of precision when casting to int maybe ?
Yes, you lose precision when messing with floats and ints.
If both floats have differing magnitude and both are using the complete precision range (of about 7 decimal digits) then yes, you will see some loss in the last places, because floats are stored in the form of (sign) (mantissa) × 2(exponent). If two values have differing exponents and you add them, then the smaller value will get reduced to less digits in the mantissa (because it has to adapt to the larger exponent):
PS> [float]([float]0.0000001 + [float]1)
1
In relation to integers, a normal 32-bit integer is capable of representing values exactly which do not fit exactly into a float. A float can still store approximately the same number, but no longer exactly. Of course, this only applies to numbers that are large enough, i. e. longer than 24 bits.Because a float has 24 bits of precision and (32-bit) integers have 32, float will still be able to retain the magnitude and most of the significant digits, but the last places may likely differ:
PS> [float]2100000050 + [float]100
2100000100
This is inherent in the use of finite-precision numerical representation schemes. Given any number that can be represented, A, there is some number that is the smallest number greater than A that can be represented, call that B. Numbers between A and B cannot be represented exactly and must be approximated.
For example, let's consider using six decimal digits because that's an easier system to understand as a starting point. If A is .333333, then B is .333334. Numbers between A and B, such 1/3, cannot be exactly represented. So if you take 1/3 and add it to itself twice (or multiply it by 3), you will get .999999, not 1. You should expect to see imprecision at the limits of the representation.
I had a problem when I was adding three floating point values and comparing them to 1.
cout << ((0.7 + 0.2 + 0.1)==1)<<endl; //output is 0
cout << ((0.7 + 0.1 + 0.2)==1)<<endl; //output is 1
Why would these values come out different?
Floating point addition is not necessarily associative. If you change the order in which you add things up, this can change the result.
The standard paper on the subject is What Every Computer Scientist Should Know about Floating Point Arithmetic. It gives the following example:
Another grey area concerns the interpretation of parentheses. Due to roundoff errors, the associative laws of algebra do not necessarily hold for floating-point numbers. For example, the expression (x+y)+z has a totally different answer than x+(y+z) when x = 1e30, y = -1e30 and z = 1 (it is 1 in the former case, 0 in the latter).
What is likely, with currently popular machines and software, is:
The compiler encoded .7 as 0x1.6666666666666p-1 (this is the hexadecimal numeral 1.6666666666666 multiplied by 2 to the power of -1), .2 as 0x1.999999999999ap-3, and .1 as 0x1.999999999999ap-4. Each of these is the number representable in floating-point that is closest to the decimal numeral you wrote.
Observe that each of these hexadecimal floating-point constants has exactly 53 bits in its significand (the "fraction" part, often inaccurately called the mantissa). The hexadecimal numeral for the significand has a "1" and thirteen more hexadecimal digits (four bits each, 52 total, 53 including the "1"), which is what the IEEE-754 standard provides for, for 64-bit binary floating-point numbers.
Let's add the numbers for .7 and .2: 0x1.6666666666666p-1 and 0x1.999999999999ap-3. First, scale the exponent of the second number to match the first. To do this, we will multiply the exponent by 4 (changing "p-3" to "p-1") and multiply the significand by 1/4, giving 0x0.66666666666668p-1. Then add 0x1.6666666666666p-1 and 0x0.66666666666668p-1, giving 0x1.ccccccccccccc8p-1. Note that this number has more than 53 bits in the significand: The "8" is the 14th digit after the period. Floating-point cannot return a result with this many bits, so it has to be rounded to the nearest representable number. In this case, there are two numbers that are equally near, 0x1.cccccccccccccp-1 and 0x1.ccccccccccccdp-1. When there is a tie, the number with a zero in the lowest bit of the significand is used. "c" is even and "d" is odd, so "c" is used. The final result of the addition is 0x1.cccccccccccccp-1.
Next, add the number for .1 (0x1.999999999999ap-4) to that. Again, we scale to make the exponents match, so 0x1.999999999999ap-4 becomes 0x.33333333333334p-1. Then add that to 0x1.cccccccccccccp-1, giving 0x1.fffffffffffff4p-1. Rounding that to 53 bits gives 0x1.fffffffffffffp-1, and that is the final result of .7+.2+.1.
Now consider .7+.1+.2. For .7+.1, add 0x1.6666666666666p-1 and 0x1.999999999999ap-4. Recall the latter is scaled to 0x.33333333333334p-1. Then the exact sum is 0x1.99999999999994p-1. Rounding that to 53 bits gives 0x1.9999999999999p-1.
Then add the number for .2 (0x1.999999999999ap-3), which is scaled to 0x0.66666666666668p-1. The exact sum is 0x2.00000000000008p-1. Floating-point significands are always scaled to start with 1 (except for special cases: zero, infinity, and very small numbers at the bottom of the representable range), so we adjust this to 0x1.00000000000004p0. Finally, we round to 53 bits, giving 0x1.0000000000000p0.
Thus, because of errors that occur when rounding, .7+.2+.1 returns 0x1.fffffffffffffp-1 (very slightly less than 1), and .7+.1+.2 returns 0x1.0000000000000p0 (exactly 1).
Floating point multiplication is not associative in C or C++.
Proof:
#include<stdio.h>
#include<time.h>
#include<stdlib.h>
using namespace std;
int main() {
int counter = 0;
srand(time(NULL));
while(counter++ < 10){
float a = rand() / 100000;
float b = rand() / 100000;
float c = rand() / 100000;
if (a*(b*c) != (a*b)*c){
printf("Not equal\n");
}
}
printf("DONE");
return 0;
}
In this program, about 30% of the time, (a*b)*c is not equal to a*(b*c).
Neither addition nor multiplication is associative with IEEE 743 double precision (64-bit) numbers. Here are examples for each (evaluated with Python 3.9.7):
>>> (.1 + .2) + .3
0.6000000000000001
>>> .1 + (.2 + .3)
0.6
>>> (.1 * .2) * .3
0.006000000000000001
>>> .1 * (.2 * .3)
0.006
Similar answer to Eric's, but for addition, and with Python.
import random
random.seed(0)
n = 1000
a = [random.random() for i in range(n)]
b = [random.random() for i in range(n)]
c = [random.random() for i in range(n)]
sum(1 if (a[i] + b[i]) + c[i] != a[i] + (b[i] + c[i]) else 0 for i in range(n))
I've read a few texts and threads showing how to convert from a decimal to IEEE 754 but I am still confused as to how I can convert the number without expanding the decimal (which is represented in scientific notation)
The number I am particularly working with is 9.07 * 10^23, but any number would do; I will figure out how to do it for my particular example.
I'm assuming you want the result to be the floating-point number closest to the decimal number, and that you are using double-precision floating-point numbers.
For most numbers, there is a way to do it relatively quickly. Here's how it works in a nutshell.
You need to split the number into either a product or a fraction of numbers that have an exact representation as a floating-point number. The largest power of 10 that is exactly representable is 10^22. So, to get 9.07e+23 in floating-point form, we can write:
9.07e+23 = 907 * 10^21
According to the IEEE-754 standard, a single floating-point operation is guaranteed to be correctly rounded, so the above product, computed as a product of 2 double precision floating-point numbers, will give the correctly rounded result.
If you were to use this in a conversion function, you would probably store the powers of 10 in an array.
Note that you can't use this method for 9.07e-23. This number equals 907 / 10^23, so the denominator would be too large to be exactly representable. In this situation, and other dealings with very large or very small numbers, you have to use some form of high-precision arithmetic.
See Fast Path Decimal to Floating-Point Conversion for further details and examples.
Converting a number from a decimal string to binary IEEE is fairly straight-forward if you know how to do IEEE floating-point addition and multiplication. (or if you're using any basic programming language like C/C++)
There's a lot of different approaches to this, but the easiest is to evaluate 9.07 * 10^23 directly.
First, start with 9.07:
9.07 = 9 + 0 * 10^-1 + 7 * 10^-2
Now evaluate 10^23. This can be done by starting with 10 and using any powering algorithm.
Then multiply the results together.
Here's a simple implementation in C/C++:
double mantissa = 9;
mantissa += 0 / 10.;
mantissa += 7 / 100.;
double exp = 1;
for (int i = 0; i < 23; i++){
exp *= 10;
}
double result = mantissa * exp;
Now, going backwards (IEEE -> to decimal) is a lot harder.
Again, there's also a lot of different approaches. Here's the easiest one I can think of it.
I'll use 1.0011101b * 2^40 as the example. (the mantissa is in binary)
First, convert the mantissa to decimal: (this should be easy, since there's no exponent)
1.0011101b * 2^40 = 1.22656 * 2^40
Now, "scale" the number such that the binary exponent vanishes. This is done by multiplying by an appropriate power of 10 to "get rid" of the binary exponent.
1.22656 * 2^40 = 1.22656 * (2^40 * 10^-12) * 10^12
= 1.22656 * (1.09951) * 10^12
= 1.34861 * 10^12
So the answer is:
1.0011101b * 2^40 = 1.34861 * 10^12
In this example, 10^12 was needed to "scale away" the 2^40. Determining the power of 10 that is needed is simply equal to:
power of 10 = (power of 2) * log(2)/log(10)
How do you print a double to a stream so that when it is read in you don't lose precision?
I tried:
std::stringstream ss;
double v = 0.1 * 0.1;
ss << std::setprecision(std::numeric_limits<T>::digits10) << v << " ";
double u;
ss >> u;
std::cout << "precision " << ((u == v) ? "retained" : "lost") << std::endl;
This did not work as I expected.
But I can increase precision (which surprised me as I thought that digits10 was the maximum required).
ss << std::setprecision(std::numeric_limits<T>::digits10 + 2) << v << " ";
// ^^^^^^ +2
It has to do with the number of significant digits and the first two don't count in (0.01).
So has anybody looked at representing floating point numbers exactly?
What is the exact magical incantation on the stream I need to do?
After some experimentation:
The trouble was with my original version. There were non-significant digits in the string after the decimal point that affected the accuracy.
So to compensate for this we can use scientific notation to compensate:
ss << std::scientific
<< std::setprecision(std::numeric_limits<double>::digits10 + 1)
<< v;
This still does not explain the need for the +1 though.
Also if I print out the number with more precision I get more precision printed out!
std::cout << std::scientific << std::setprecision(std::numeric_limits<double>::digits10) << v << "\n";
std::cout << std::scientific << std::setprecision(std::numeric_limits<double>::digits10 + 1) << v << "\n";
std::cout << std::scientific << std::setprecision(std::numeric_limits<double>::digits) << v << "\n";
It results in:
1.000000000000000e-02
1.0000000000000002e-02
1.00000000000000019428902930940239457413554200000000000e-02
Based on #Stephen Canon answer below:
We can print out exactly by using the printf() formatter, "%a" or "%A". To achieve this in C++ we need to use the fixed and scientific manipulators (see n3225: 22.4.2.2.2p5 Table 88)
std::cout.flags(std::ios_base::fixed | std::ios_base::scientific);
std::cout << v;
For now I have defined:
template<typename T>
std::ostream& precise(std::ostream& stream)
{
std::cout.flags(std::ios_base::fixed | std::ios_base::scientific);
return stream;
}
std::ostream& preciselngd(std::ostream& stream){ return precise<long double>(stream);}
std::ostream& precisedbl(std::ostream& stream) { return precise<double>(stream);}
std::ostream& preciseflt(std::ostream& stream) { return precise<float>(stream);}
Next: How do we handle NaN/Inf?
It's not correct to say "floating point is inaccurate", although I admit that's a useful simplification. If we used base 8 or 16 in real life then people around here would be saying "base 10 decimal fraction packages are inaccurate, why did anyone ever cook those up?".
The problem is that integral values translate exactly from one base into another, but fractional values do not, because they represent fractions of the integral step and only a few of them are used.
Floating point arithmetic is technically perfectly accurate. Every calculation has one and only one possible result. There is a problem, and it is that most decimal fractions have base-2 representations that repeat. In fact, in the sequence 0.01, 0.02, ... 0.99, only a mere 3 values have exact binary representations. (0.25, 0.50, and 0.75.) There are 96 values that repeat and therefore are obviously not represented exactly.
Now, there are a number of ways to write and read back floating point numbers without losing a single bit. The idea is to avoid trying to express the binary number with a base 10 fraction.
Write them as binary. These days, everyone implements the IEEE-754 format so as long as you choose a byte order and write or read only that byte order, then the numbers will be portable.
Write them as 64-bit integer values. Here you can use the usual base 10. (Because you are representing the 64-bit aliased integer, not the 52-bit fraction.)
You can also just write more decimal fraction digits. Whether this is bit-for-bit accurate will depend on the quality of the conversion libraries and I'm not sure I would count on perfect accuracy (from the software) here. But any errors will be exceedingly small and your original data certainly has no information in the low bits. (None of the constants of physics and chemistry are known to 52 bits, nor has any distance on earth ever been measured to 52 bits of precision.) But for a backup or restore where bit-for-bit accuracy might be compared automatically, this obviously isn't ideal.
Don't print floating-point values in decimal if you don't want to lose precision. Even if you print enough digits to represent the number exactly, not all implementations have correctly-rounded conversions to/from decimal strings over the entire floating-point range, so you may still lose precision.
Use hexadecimal floating point instead. In C:
printf("%a\n", yourNumber);
C++0x provides the hexfloat manipulator for iostreams that does the same thing (on some platforms, using the std::hex modifier has the same result, but this is not a portable assumption).
Using hex floating point is preferred for several reasons.
First, the printed value is always exact. No rounding occurs in writing or reading a value formatted in this way. Beyond the accuracy benefits, this means that reading and writing such values can be faster with a well tuned I/O library. They also require fewer digits to represent values exactly.
I got interested in this question because I'm trying to (de)serialize my data to & from JSON.
I think I have a clearer explanation (with less hand waiving) for why 17 decimal digits are sufficient to reconstruct the original number losslessly:
Imagine 3 number lines:
1. for the original base 2 number
2. for the rounded base 10 representation
3. for the reconstructed number (same as #1 because both in base 2)
When you convert to base 10, graphically, you choose the tic on the 2nd number line closest to the tic on the 1st. Likewise when you reconstruct the original from the rounded base 10 value.
The critical observation I had was that in order to allow exact reconstruction, the base 10 step size (quantum) has to be < the base 2 quantum. Otherwise, you inevitably get the bad reconstruction shown in red.
Take the specific case of when the exponent is 0 for the base2 representation. Then the base2 quantum will be 2^-52 ~= 2.22 * 10^-16. The closest base 10 quantum that's less than this is 10^-16. Now that we know the required base 10 quantum, how many digits will be needed to encode all possible values? Given that we're only considering the case of exponent = 0, the dynamic range of values we need to represent is [1.0, 2.0). Therefore, 17 digits would be required (16 digits for fraction and 1 digit for integer part).
For exponents other than 0, we can use the same logic:
exponent base2 quant. base10 quant. dynamic range digits needed
---------------------------------------------------------------------
1 2^-51 10^-16 [2, 4) 17
2 2^-50 10^-16 [4, 8) 17
3 2^-49 10^-15 [8, 16) 17
...
32 2^-20 10^-7 [2^32, 2^33) 17
1022 9.98e291 1.0e291 [4.49e307,8.99e307) 17
While not exhaustive, the table shows the trend that 17 digits are sufficient.
Hope you like my explanation.
In C++20 you'll be able to use std::format to do this:
std::stringstream ss;
double v = 0.1 * 0.1;
ss << std::format("{}", v);
double u;
ss >> u;
assert(v == u);
The default floating-point format is the shortest decimal representation with a round-trip guarantee. The advantage of this method compared to using the precision of max_digits10 (not digits10 which is not suitable for round trip through decimal) from std::numeric_limits is that it doesn't print unnecessary digits.
In the meantime you can use the {fmt} library, std::format is based on. For example (godbolt):
fmt::print("{}", 0.1 * 0.1);
Output (assuming IEEE754 double):
0.010000000000000002
{fmt} uses the Dragonbox algorithm for fast binary floating point to decimal conversion. In addition to giving the shortest representation it is 20-30x faster than common standard library implementations of printf and iostreams.
Disclaimer: I'm the author of {fmt} and C++20 std::format.
A double has the precision of 52 binary digits or 15.95 decimal digits. See http://en.wikipedia.org/wiki/IEEE_754-2008. You need at least 16 decimal digits to record the full precision of a double in all cases. [But see fourth edit, below].
By the way, this means significant digits.
Answer to OP edits:
Your floating point to decimal string runtime is outputing way more digits than are significant. A double can only hold 52 bits of significand (actually, 53, if you count a "hidden" 1 that is not stored). That means the the resolution is not more than 2 ^ -53 = 1.11e-16.
For example: 1 + 2 ^ -52 = 1.0000000000000002220446049250313 . . . .
Those decimal digits, .0000000000000002220446049250313 . . . . are the smallest binary "step" in a double when converted to decimal.
The "step" inside the double is:
.0000000000000000000000000000000000000000000000000001 in binary.
Note that the binary step is exact, while the decimal step is inexact.
Hence the decimal representation above,
1.0000000000000002220446049250313 . . .
is an inexact representation of the exact binary number:
1.0000000000000000000000000000000000000000000000000001.
Third Edit:
The next possible value for a double, which in exact binary is:
1.0000000000000000000000000000000000000000000000000010
converts inexactly in decimal to
1.0000000000000004440892098500626 . . . .
So all of those extra digits in the decimal are not really significant, they are just base conversion artifacts.
Fourth Edit:
Though a double stores at most 16 significant decimal digits, sometimes 17 decimal digits are necessary to represent the number. The reason has to do with digit slicing.
As I mentioned above, there are 52 + 1 binary digits in the double. The "+ 1" is an assumed leading 1, and is neither stored nor significant. In the case of an integer, those 52 binary digits form a number between 0 and 2^53 - 1. How many decimal digits are necessary to store such a number? Well, log_10 (2^53 - 1) is about 15.95. So at most 16 decimal digits are necessary. Let's label these d_0 to d_15.
Now consider that IEEE floating point numbers also have an binary exponent. What happens when we increment the exponet by, say, 2? We have multiplied our 52-bit number, whatever it was, by 4. Now, instead of our 52 binary digits aligning perfectly with our decimal digits d_0 to d_15, we have some significant binary digits represented in d_16. However, since we multiplied by something less than 10, we still have significant binary digits represented in d_0. So our 15.95 decimal digits now occuply d_1 to d_15, plus some upper bits of d_0 and some lower bits of d_16. This is why 17 decimal digits are sometimes needed to represent a IEEE double.
Fifth Edit
Fixed numerical errors
The easiest way (for IEEE 754 double) to guarantee a round-trip conversion is to always use 17 significant digits. But that has the disadvantage of sometimes including unnecessary noise digits (0.1 → "0.10000000000000001").
An approach that's worked for me is to sprintf the number with 15 digits of precision, then check if atof gives you back the original value. If it doesn't, try 16 digits. If that doesn't work, use 17.
You might want to try David Gay's algorithm (used in Python 3.1 to implement float.__repr__).
Thanks to ThomasMcLeod for pointing out the error in my table computation
To guarantee round-trip conversion using 15 or 16 or 17 digits is only possible for a comparatively few cases. The number 15.95 comes from taking 2^53 (1 implicit bit + 52 bits in the significand/"mantissa") which comes out to an integer in the range 10^15 to 10^16 (closer to 10^16).
Consider a double precision value x with an exponent of 0, i.e. it falls into the floating point range range 1.0 <= x < 2.0. The implicit bit will mark the 2^0 component (part) of x. The highest explicit bit of the significand will denote the next lower exponent (from 0) <=> -1 => 2^-1 or the 0.5 component.
The next bit 0.25, the ones after 0.125, 0.0625, 0.03125, 0.015625 and so on (see table below). The value 1.5 will thus be represented by two components added together: the implicit bit denoting 1.0 and the highest explicit significand bit denoting 0.5.
This illustrates that from the implicit bit downward you have 52 additional, explicit bits to represent possible components where the smallest is 0 (exponent) - 52 (explicit bits in significand) = -52 => 2^-52 which according to the table below is ... well you can see for yourselves that it comes out to quite a bit more than 15.95 significant digits (37 to be exact). To put it another way the smallest number in the 2^0 range that is != 1.0 itself is 2^0+2^-52 which is 1.0 + the number next to 2^-52 (below) = (exactly) 1.0000000000000002220446049250313080847263336181640625, a value which I count as being 53 significant digits long. With 17 digit formatting "precision" the number will display as 1.0000000000000002 and this would depend on the library converting correctly.
So maybe "round-trip conversion in 17 digits" is not really a concept that is valid (enough).
2^ -1 = 0.5000000000000000000000000000000000000000000000000000
2^ -2 = 0.2500000000000000000000000000000000000000000000000000
2^ -3 = 0.1250000000000000000000000000000000000000000000000000
2^ -4 = 0.0625000000000000000000000000000000000000000000000000
2^ -5 = 0.0312500000000000000000000000000000000000000000000000
2^ -6 = 0.0156250000000000000000000000000000000000000000000000
2^ -7 = 0.0078125000000000000000000000000000000000000000000000
2^ -8 = 0.0039062500000000000000000000000000000000000000000000
2^ -9 = 0.0019531250000000000000000000000000000000000000000000
2^-10 = 0.0009765625000000000000000000000000000000000000000000
2^-11 = 0.0004882812500000000000000000000000000000000000000000
2^-12 = 0.0002441406250000000000000000000000000000000000000000
2^-13 = 0.0001220703125000000000000000000000000000000000000000
2^-14 = 0.0000610351562500000000000000000000000000000000000000
2^-15 = 0.0000305175781250000000000000000000000000000000000000
2^-16 = 0.0000152587890625000000000000000000000000000000000000
2^-17 = 0.0000076293945312500000000000000000000000000000000000
2^-18 = 0.0000038146972656250000000000000000000000000000000000
2^-19 = 0.0000019073486328125000000000000000000000000000000000
2^-20 = 0.0000009536743164062500000000000000000000000000000000
2^-21 = 0.0000004768371582031250000000000000000000000000000000
2^-22 = 0.0000002384185791015625000000000000000000000000000000
2^-23 = 0.0000001192092895507812500000000000000000000000000000
2^-24 = 0.0000000596046447753906250000000000000000000000000000
2^-25 = 0.0000000298023223876953125000000000000000000000000000
2^-26 = 0.0000000149011611938476562500000000000000000000000000
2^-27 = 0.0000000074505805969238281250000000000000000000000000
2^-28 = 0.0000000037252902984619140625000000000000000000000000
2^-29 = 0.0000000018626451492309570312500000000000000000000000
2^-30 = 0.0000000009313225746154785156250000000000000000000000
2^-31 = 0.0000000004656612873077392578125000000000000000000000
2^-32 = 0.0000000002328306436538696289062500000000000000000000
2^-33 = 0.0000000001164153218269348144531250000000000000000000
2^-34 = 0.0000000000582076609134674072265625000000000000000000
2^-35 = 0.0000000000291038304567337036132812500000000000000000
2^-36 = 0.0000000000145519152283668518066406250000000000000000
2^-37 = 0.0000000000072759576141834259033203125000000000000000
2^-38 = 0.0000000000036379788070917129516601562500000000000000
2^-39 = 0.0000000000018189894035458564758300781250000000000000
2^-40 = 0.0000000000009094947017729282379150390625000000000000
2^-41 = 0.0000000000004547473508864641189575195312500000000000
2^-42 = 0.0000000000002273736754432320594787597656250000000000
2^-43 = 0.0000000000001136868377216160297393798828125000000000
2^-44 = 0.0000000000000568434188608080148696899414062500000000
2^-45 = 0.0000000000000284217094304040074348449707031250000000
2^-46 = 0.0000000000000142108547152020037174224853515625000000
2^-47 = 0.0000000000000071054273576010018587112426757812500000
2^-48 = 0.0000000000000035527136788005009293556213378906250000
2^-49 = 0.0000000000000017763568394002504646778106689453125000
2^-50 = 0.0000000000000008881784197001252323389053344726562500
2^-51 = 0.0000000000000004440892098500626161694526672363281250
2^-52 = 0.0000000000000002220446049250313080847263336181640625
#ThomasMcLeod: I think the significant digit rule comes from my field, physics, and means something more subtle:
If you have a measurement that gets you the value 1.52 and you cannot read any more detail off the scale, and say you are supposed to add another number (for example of another measurement because this one's scale was too small) to it, say 2, then the result (obviously) has only 2 decimal places, i.e. 3.52.
But likewise, if you add 1.1111111111 to the value 1.52, you get the value 2.63 (and nothing more!).
The reason for the rule is to prevent you from kidding yourself into thinking you got more information out of a calculation than you put in by the measurement (which is impossible, but would seem that way by filling it with garbage, see above).
That said, this specific rule is for addition only (for addition: the error of the result is the sum of the two errors - so if you measure just one badly, though luck, there goes your precision...).
How to get the other rules:
Let's say a is the measured number and δa the error. Let's say your original formula was:
f:=m a
Let's say you also measure m with error δm (let that be the positive side).
Then the actual limit is:
f_up=(m+δm) (a+δa)
and
f_down=(m-δm) (a-δa)
So,
f_up =m a+δm δa+(δm a+m δa)
f_down=m a+δm δa-(δm a+m δa)
Hence, now the significant digits are even less:
f_up ~m a+(δm a+m δa)
f_down~m a-(δm a+m δa)
and so
δf=δm a+m δa
If you look at the relative error, you get:
δf/f=δm/m+δa/a
For division it is
δf/f=δm/m-δa/a
Hope that gets the gist across and hope I didn't make too many mistakes, it's late here :-)
tl,dr: Significant digits mean how many of the digits in the output actually come from the digits in your input (in the real world, not the distorted picture that floating point numbers have).
If your measurements were 1 with "no" error and 3 with "no" error and the function is supposed to be 1/3, then yes, all infinite digits are actual significant digits. Otherwise, the inverse operation would not work, so obviously they have to be.
If significant digit rule means something completely different in another field, carry on :-)