Algorithm for Simplifying Decimal to Fractions
I Tried Writing an Algorithm to Simplify a Decimal to a Fraction and Realized It Wasn't Too Simple. Write 0.333333. .. as 1/3 for Example. or 0.1666667, Which...
I tried writing an algorithm to simplify a decimal to a fraction and realized it wasn't too simple.
Write 0.333333... as 1/3 for example.
Or 0.1666667, which is 1/6.
Surprisingly I looked online and all the code I found was either too long, or wouldn't work in some cases. What was even more annoying was that they didn't work for recurring decimals.
How can one simplify a decimal to a fraction?
25 Answers
The algorithm that the other people have given you gets the answer by calculating the Continued Fraction of the number. This gives a fractional sequence which is guaranteed to converge very, very rapidly. However it is not guaranteed to give you the smallest fraction that is within a distance epsilon of a real number. To find that you have to walk the Stern-Brocot tree.
To do that you subtract off the floor to get the number in the range [0, 1), then your lower estimate is 0, and your upper estimate is 1. Now do a binary search until you are close enough. At each iteration if your lower is a/b and your upper is c/d your middle is (a+c)/(b+d). Test your middle against x, and either make the middle the upper, the lower, or return your final answer.
Here is some very non-idiomatic (and hence, hopefully, readable even if you don't know the language) Python that implements this algorithm.
def float_to_fraction (x, error=0.000001):
n = int(math.floor(x))
x -= n
if x < error:
return (n, 1)
elif 1 - error < x:
return (n+1, 1)
# The lower fraction is 0/1
lower_n = 0
lower_d = 1
# The upper fraction is 1/1
upper_n = 1
upper_d = 1
while True:
# The middle fraction is (lower_n + upper_n) / (lower_d + upper_d)
middle_n = lower_n + upper_n
middle_d = lower_d + upper_d
# If x + error < middle
if middle_d * (x + error) < middle_n:
# middle is our new upper
upper_n = middle_n
upper_d = middle_d
# Else If middle < x - error
elif middle_n < (x - error) * middle_d:
# middle is our new lower
lower_n = middle_n
lower_d = middle_d
# Else middle is our best fraction
else:
return (n * middle_d + middle_n, middle_d)
(code improved Feb 2017 - scroll down to 'optimization'...)
(algorithm comparison table at the end of this answer)
I implemented btilly's answer in C# and...
- added support for negative numbers
- provide an
accuracyparameter to specify the max. relative error, not the max. absolute error;0.01would find a fraction within 1% of the value. - provide an optimization
Double.NaNandDouble.Infinityare not supported; you might want to handle those (example here).
public Fraction RealToFraction(double value, double accuracy)
{
if (accuracy <= 0.0 || accuracy >= 1.0)
{
throw new ArgumentOutOfRangeException("accuracy", "Must be > 0 and < 1.");
}
int sign = Math.Sign(value);
if (sign == -1)
{
value = Math.Abs(value);
}
// Accuracy is the maximum relative error; convert to absolute maxError
double maxError = sign == 0 ? accuracy : value * accuracy;
int n = (int) Math.Floor(value);
value -= n;
if (value < maxError)
{
return new Fraction(sign * n, 1);
}
if (1 - maxError < value)
{
return new Fraction(sign * (n + 1), 1);
}
// The lower fraction is 0/1
int lower_n = 0;
int lower_d = 1;
// The upper fraction is 1/1
int upper_n = 1;
int upper_d = 1;
while (true)
{
// The middle fraction is (lower_n + upper_n) / (lower_d + upper_d)
int middle_n = lower_n + upper_n;
int middle_d = lower_d + upper_d;
if (middle_d * (value + maxError) < middle_n)
{
// real + error < middle : middle is our new upper
upper_n = middle_n;
upper_d = middle_d;
}
else if (middle_n < (value - maxError) * middle_d)
{
// middle < real - error : middle is our new lower
lower_n = middle_n;
lower_d = middle_d;
}
else
{
// Middle is our best fraction
return new Fraction((n * middle_d + middle_n) * sign, middle_d);
}
}
}
The Fraction type is just a simple struct. Of course, use your own preferred type... (I like this one by Rick Davin.)
public struct Fraction
{
public Fraction(int n, int d)
{
N = n;
D = d;
}
public int N { get; private set; }
public int D { get; private set; }
}
Feb 2017 optimization
For certain values, like 0.01, 0.001, etc. the algorithm goes through hundreds or thousands of linear iterations. To fix this, I implemented a binary way of finding the final value -- thanks to btilly for this idea. Inside the if-statement substitute the following:
// real + error < middle : middle is our new upper
Seek(ref upper_n, ref upper_d, lower_n, lower_d, (un, ud) => (lower_d + ud) * (value + maxError) < (lower_n + un));
and
// middle < real - error : middle is our new lower
Seek(ref lower_n, ref lower_d, upper_n, upper_d, (ln, ld) => (ln + upper_n) < (value - maxError) * (ld + upper_d));
Here is the Seek method implementation:
/// <summary>
/// Binary seek for the value where f() becomes false.
/// </summary>
void Seek(ref int a, ref int b, int ainc, int binc, Func<int, int, bool> f)
{
a += ainc;
b += binc;
if (f(a, b))
{
int weight = 1;
do
{
weight *= 2;
a += ainc * weight;
b += binc * weight;
}
while (f(a, b));
do
{
weight /= 2;
int adec = ainc * weight;
int bdec = binc * weight;
if (!f(a - adec, b - bdec))
{
a -= adec;
b -= bdec;
}
}
while (weight > 1);
}
}
Algorithm comparison table
You may want to copy the table to your text editor for full screen viewing.
Accuracy: 1.0E-3 | Stern-Brocot OPTIMIZED | Eppstein | Richards
Input | Result Error Iterations Iterations | Result Error Iterations | Result Error Iterations
======================| =====================================================| =========================================| =========================================
0 | 0/1 (zero) 0 0 0 | 0/1 (zero) 0 0 | 0/1 (zero) 0 0
1 | 1/1 0 0 0 | 1001/1000 1.0E-3 1 | 1/1 0 0
3 | 3/1 0 0 0 | 1003/334 1.0E-3 1 | 3/1 0 0
-1 | -1/1 0 0 0 | -1001/1000 1.0E-3 1 | -1/1 0 0
-3 | -3/1 0 0 0 | -1003/334 1.0E-3 1 | -3/1 0 0
0.999999 | 1/1 1.0E-6 0 0 | 1000/1001 -1.0E-3 2 | 1/1 1.0E-6 0
-0.999999 | -1/1 1.0E-6 0 0 | -1000/1001 -1.0E-3 2 | -1/1 1.0E-6 0
1.000001 | 1/1 -1.0E-6 0 0 | 1001/1000 1.0E-3 1 | 1/1 -1.0E-6 0
-1.000001 | -1/1 -1.0E-6 0 0 | -1001/1000 1.0E-3 1 | -1/1 -1.0E-6 0
0.50 (1/2) | 1/2 0 1 1 | 999/1999 -5.0E-4 2 | 1/2 0 1
0.33... (1/3) | 1/3 0 2 2 | 999/2998 -3.3E-4 2 | 1/3 0 1
0.67... (2/3) | 2/3 0 2 2 | 999/1498 3.3E-4 3 | 2/3 0 2
0.25 (1/4) | 1/4 0 3 3 | 999/3997 -2.5E-4 2 | 1/4 0 1
0.11... (1/9) | 1/9 0 8 4 | 999/8992 -1.1E-4 2 | 1/9 0 1
0.09... (1/11) | 1/11 0 10 5 | 999/10990 -9.1E-5 2 | 1/11 0 1
0.62... (307/499) | 8/13 2.5E-4 5 5 | 913/1484 -2.2E-6 8 | 8/13 2.5E-4 5
0.14... (33/229) | 15/104 8.7E-4 20 9 | 974/6759 -4.5E-6 6 | 16/111 2.7E-4 3
0.05... (33/683) | 7/145 -8.4E-4 24 10 | 980/20283 1.5E-6 7 | 10/207 -1.5E-4 4
0.18... (100/541) | 17/92 -3.3E-4 11 10 | 939/5080 -2.0E-6 8 | 17/92 -3.3E-4 4
0.06... (33/541) | 5/82 -3.7E-4 19 8 | 995/16312 -1.9E-6 6 | 5/82 -3.7E-4 4
0.1 | 1/10 0 9 5 | 999/9991 -1.0E-4 2 | 1/10 0 1
0.2 | 1/5 0 4 3 | 999/4996 -2.0E-4 2 | 1/5 0 1
0.3 | 3/10 0 5 5 | 998/3327 -1.0E-4 4 | 3/10 0 3
0.4 | 2/5 0 3 3 | 999/2497 2.0E-4 3 | 2/5 0 2
0.5 | 1/2 0 1 1 | 999/1999 -5.0E-4 2 | 1/2 0 1
0.6 | 3/5 0 3 3 | 1000/1667 -2.0E-4 4 | 3/5 0 3
0.7 | 7/10 0 5 5 | 996/1423 -1.0E-4 4 | 7/10 0 3
0.8 | 4/5 0 4 3 | 997/1246 2.0E-4 3 | 4/5 0 2
0.9 | 9/10 0 9 5 | 998/1109 -1.0E-4 4 | 9/10 0 3
0.01 | 1/100 0 99 8 | 999/99901 -1.0E-5 2 | 1/100 0 1
0.001 | 1/1000 0 999 11 | 999/999001 -1.0E-6 2 | 1/1000 0 1
0.0001 | 1/9991 9.0E-4 9990 15 | 999/9990001 -1.0E-7 2 | 1/10000 0 1
1E-05 | 1/99901 9.9E-4 99900 18 | 1000/99999999 1.0E-8 3 | 1/99999 1.0E-5 1
0.33333333333 | 1/3 1.0E-11 2 2 | 1000/3001 -3.3E-4 2 | 1/3 1.0E-11 1
0.3 | 3/10 0 5 5 | 998/3327 -1.0E-4 4 | 3/10 0 3
0.33 | 30/91 -1.0E-3 32 8 | 991/3003 1.0E-5 3 | 33/100 0 2
0.333 | 167/502 -9.9E-4 169 11 | 1000/3003 1.0E-6 3 | 333/1000 0 2
0.7777 | 7/9 1.0E-4 5 4 | 997/1282 -1.1E-5 4 | 7/9 1.0E-4 3
0.101 | 10/99 1.0E-4 18 10 | 919/9099 1.1E-6 5 | 10/99 1.0E-4 3
0.10001 | 1/10 -1.0E-4 9 5 | 1/10 -1.0E-4 4 | 1/10 -1.0E-4 2
0.100000001 | 1/10 -1.0E-8 9 5 | 1000/9999 1.0E-4 3 | 1/10 -1.0E-8 2
0.001001 | 1/999 1.0E-6 998 11 | 1/999 1.0E-6 3 | 1/999 1.0E-6 1
0.0010000001 | 1/1000 -1.0E-7 999 11 | 1000/999999 9.0E-7 3 | 1/1000 -1.0E-7 2
0.11 | 10/91 -1.0E-3 18 9 | 1000/9091 -1.0E-5 4 | 10/91 -1.0E-3 2
0.1111 | 1/9 1.0E-4 8 4 | 1000/9001 -1.1E-5 2 | 1/9 1.0E-4 1
0.111111111111 | 1/9 1.0E-12 8 4 | 1000/9001 -1.1E-4 2 | 1/9 1.0E-12 1
1 | 1/1 0 0 0 | 1001/1000 1.0E-3 1 | 1/1 0 0
-1 | -1/1 0 0 0 | -1001/1000 1.0E-3 1 | -1/1 0 0
-0.5 | -1/2 0 1 1 | -999/1999 -5.0E-4 2 | -1/2 0 1
3.14 | 22/7 9.1E-4 6 4 | 964/307 2.1E-5 3 | 22/7 9.1E-4 1
3.1416 | 22/7 4.0E-4 6 4 | 732/233 9.8E-6 3 | 22/7 4.0E-4 1
3.14... (pi) | 22/7 4.0E-4 6 4 | 688/219 -1.3E-5 4 | 22/7 4.0E-4 1
0.14 | 7/50 0 13 7 | 995/7107 2.0E-5 3 | 7/50 0 2
0.1416 | 15/106 -6.4E-4 21 8 | 869/6137 9.2E-7 5 | 16/113 -5.0E-5 2
2.72... (e) | 68/25 6.3E-4 7 7 | 878/323 -5.7E-6 8 | 87/32 1.7E-4 5
0.141592653589793 | 15/106 -5.9E-4 21 8 | 991/6999 -7.0E-6 4 | 15/106 -5.9E-4 2
-1.33333333333333 | -4/3 2.5E-15 2 2 | -1001/751 -3.3E-4 2 | -4/3 2.5E-15 1
-1.3 | -13/10 0 5 5 | -992/763 1.0E-4 3 | -13/10 0 2
-1.33 | -97/73 -9.3E-4 26 8 | -935/703 1.1E-5 3 | -133/100 0 2
-1.333 | -4/3 2.5E-4 2 2 | -1001/751 -8.3E-5 2 | -4/3 2.5E-4 1
-1.33333337 | -4/3 -2.7E-8 2 2 | -999/749 3.3E-4 3 | -4/3 -2.7E-8 2
-1.7 | -17/10 0 5 5 | -991/583 -1.0E-4 4 | -17/10 0 3
-1.37 | -37/27 2.7E-4 7 7 | -996/727 1.0E-5 7 | -37/27 2.7E-4 5
-1.33337 | -4/3 -2.7E-5 2 2 | -999/749 3.1E-4 3 | -4/3 -2.7E-5 2
0.047619 | 1/21 1.0E-6 20 6 | 1000/21001 -4.7E-5 2 | 1/21 1.0E-6 1
12.125 | 97/8 0 7 4 | 982/81 -1.3E-4 2 | 97/8 0 1
5.5 | 11/2 0 1 1 | 995/181 -5.0E-4 2 | 11/2 0 1
0.1233333333333 | 9/73 -3.7E-4 16 8 | 971/7873 -3.4E-6 4 | 9/73 -3.7E-4 2
0.7454545454545 | 38/51 -4.8E-4 15 8 | 981/1316 -1.9E-5 6 | 38/51 -4.8E-4 4
0.01024801004 | 2/195 8.2E-4 98 9 | 488/47619 2.0E-8 13 | 2/195 8.2E-4 3
0.99011 | 91/92 -9.9E-4 91 8 | 801/809 1.3E-6 5 | 100/101 -1.1E-5 2
0.9901134545 | 91/92 -9.9E-4 91 8 | 601/607 1.9E-6 5 | 100/101 -1.5E-5 2
0.19999999 | 1/5 5.0E-8 4 3 | 1000/5001 -2.0E-4 2 | 1/5 5.0E-8 1
0.20000001 | 1/5 -5.0E-8 4 3 | 1000/4999 2.0E-4 3 | 1/5 -5.0E-8 2
5.0183168565E-05 | 1/19908 9.5E-4 19907 16 | 1000/19927001 -5.0E-8 2 | 1/19927 5.2E-12 1
3.909E-07 | 1/2555644 1.0E-3 2555643 23 | 1/1 2.6E6 (!) 1 | 1/2558199 1.1E-8 1
88900003.001 |88900003/1 -1.1E-11 0 0 |88900004/1 1.1E-8 1 |88900003/1 -1.1E-11 0
0.26... (5/19) | 5/19 0 7 6 | 996/3785 -5.3E-5 4 | 5/19 0 3
0.61... (37/61) | 17/28 9.7E-4 8 7 | 982/1619 -1.7E-5 8 | 17/28 9.7E-4 5
| | |
Accuracy: 1.0E-4 | Stern-Brocot OPTIMIZED | Eppstein | Richards
Input | Result Error Iterations Iterations | Result Error Iterations | Result Error Iterations
======================| =====================================================| =========================================| =========================================
0.62... (307/499) | 227/369 -8.8E-5 33 11 | 9816/15955 -2.0E-7 8 | 299/486 -6.7E-6 6
0.05... (33/683) | 23/476 6.4E-5 27 12 | 9989/206742 1.5E-7 7 | 23/476 6.4E-5 5
0.06... (33/541) | 28/459 6.6E-5 24 12 | 9971/163464 -1.9E-7 6 | 33/541 0 5
1E-05 | 1/99991 9.0E-5 99990 18 | 10000/999999999 1.0E-9 3 | 1/99999 1.0E-5 1
0.333 | 303/910 -9.9E-5 305 12 | 9991/30003 1.0E-7 3 | 333/1000 0 2
0.7777 | 556/715 -1.0E-4 84 12 | 7777/10000 0 8 | 1109/1426 -1.8E-7 4
3.14... (pi) | 289/92 -9.2E-5 19 8 | 9918/3157 -8.1E-7 4 | 333/106 -2.6E-5 2
2.72... (e) | 193/71 1.0E-5 10 9 | 9620/3539 6.3E-8 11 | 193/71 1.0E-5 7
0.7454545454545 | 41/55 6.1E-14 16 8 | 9960/13361 -1.8E-6 6 | 41/55 6.1E-14 5
0.01024801004 | 7/683 8.7E-5 101 12 | 9253/902907 -1.3E-10 16 | 7/683 8.7E-5 5
0.99011 | 100/101 -1.1E-5 100 8 | 901/910 -1.1E-7 6 | 100/101 -1.1E-5 2
0.9901134545 | 100/101 -1.5E-5 100 8 | 8813/8901 1.6E-8 7 | 100/101 -1.5E-5 2
0.26... (5/19) | 5/19 0 7 6 | 9996/37985 -5.3E-6 4 | 5/19 0 3
0.61... (37/61) | 37/61 0 10 8 | 9973/16442 -1.6E-6 8 | 37/61 0 7
Performance comparison
I performed detailed speed tests and plotted the results. Not looking at quality and only speed:
- The Stern-Brocot optimization slows it down by at most a factor 2, but the original Stern-Brocot can be hundreds or thousands times slower when it hits the unlucky values mentioned. That's still only a couple of microseconds though per call.
- Richards is consistently fast.
- Eppstein is around 3 times slower than the others.
Stern-Brocot and Richards compared:
- Both return nice fractions.
- Richards often results in a smaller error. It is also a bit faster.
- Stern-Brocot walks down the S-B tree. It finds the fraction of the lowest denominator that meets the required accuracy, then stops.
If you do not require the lowest denominator fraction, Richards is a good choice.
I know you said you searched online, but if you missed the following paper it might be of some help. It includes a code example in Pascal.
Algorithm To Convert A Decimal To A Fraction*
Alternatively, as part of it's standard library, Ruby has code that deals with rational numbers. It can convert from floats to rationals and vice versa. I believe you can look through the code as well. The documentation is found here. I know you're not using Ruby, but it might help to look at the algorithms.
Additionally, you can call Ruby code from C# (or even write Ruby code inside a C# code file) if you use IronRuby, which runs on top of the .net framework.
*Updated to a new link as it appears the original URL is broken ()
The most populair solutions to this problem are Richards’ algorithm and the Stern-Brocot algorithm, implemented by btilly with speed optimalization by btilly and Jay Zed. Richards’ algorithm is the fastest, but does not guarantee to return the best fraction.
I have an solution to this problem which always gives the best fraction and is also faster than all of the algorithms above. Here is the algorithm in C# (explanation and speed test below).
This is a short algorithm without comments. An complete version is provided in the source code at the end.
public static Fraction DoubleToFractionSjaak(double value, double accuracy)
{
int sign = value < 0 ? -1 : 1;
value = value < 0 ? -value : value;
int integerpart = (int)value;
value -= integerpart;
double minimalvalue = value - accuracy;
if (minimalvalue < 0.0) return new Fraction(sign * integerpart, 1);
double maximumvalue = value + accuracy;
if (maximumvalue > 1.0) return new Fraction(sign * (integerpart + 1), 1);
int a = 0;
int b = 1;
int c = 1;
int d = (int)(1 / maximumvalue);
while (true)
{
int n = (int)((b * minimalvalue - a) / (c - d * minimalvalue));
if (n == 0) break;
a += n * c;
b += n * d;
n = (int)((c - d * maximumvalue) / (b * maximumvalue - a));
if (n == 0) break;
c += n * a;
d += n * b;
}
int denominator = b + d;
return new Fraction(sign * (integerpart * denominator + (a + c)), denominator);
}
Where Fraction is a simple class to store a fraction, like the following:
public class Fraction
{
public int Numerator { get; private set; }
public int Denominator { get; private set; }
public Fraction(int numerator, int denominator)
{
Numerator = numerator;
Denominator = denominator;
}
}
Must Read
How it works
Like the other solutions mentioned, my solution is based on continued fraction. Other solutions like the one from Eppstein or solutions based on repeating decimals proved to be slower and/or give suboptimal results.
Continued fraction
Solutions based on continued fraction are mostly based on two algorithms, both described in an article by Ian Richards published here in 1981. He called them the “slow continued fraction algorithm” and the “fast continued fraction algorithm”. The first is known as the the Stern-Brocot algorithm while the latter is known as Richards’ algorithm.
My algorithm (short explanation)
To fully understand my algorithm, you need to have read the article by Ian Richards or at least understand what a Farey pair is. Furthermore, read the algorithm with comments at the end of this article.
The algorithm is using a Farey pair, containing a left and a right fraction. By repeatedly taking the mediant it is closing in on the target value. This is just like the slow algorithm but there are two major differences:
- Multiple iterations are performed at once as long as the mediant stay on one side of the target value.
- The left and right fraction cannot come closer to the target value than the given accuracy.
Alternately the right and left side of the target value are checked. If the algorithm cannot produce a result closer to the target value, the process ends. The resulting mediant is the optimal solution.
Maya is a hardware enthusiast who tests and reviews smart home devices, smartphones, wearables, and audio gear. She focuses on practical consumer value and build quality.
