How to compare that sequence of doubles are all “a

Question

I have a method in java that returns a double number and I want to compare every double number that is returned every time I call the method(say 5 times), so that I can conclude that the number returned is almost the same every time.

How can I do this?

Answer 1

It depends on what you mean by similar. If you want to compare two numbers within an absolute error e.g. 1e-6 you can use epsilon. If you want to compare two double regardless of scale. e.g. 1.1e-20 and 1.3e-20 are not similar but 1.1e20 and 1.1e20+1e5 are you can compare the raw value.

public static void main(String... args) throws IOException {
    test(1.1e-20, 1.3e-20);
    test(1.1e20, 1.1e20 + 1e5);
}

private static void test(double a, double b) {
    System.out.println(a + " and " + b + ", similar= " + similarUnscaled(a, b, 10));
}

public static boolean similarUnscaled(double a, double b, long representationDifference) {
    long a2 = Double.doubleToRawLongBits(a);
    long b2 = Double.doubleToRawLongBits(b);
    // avoid overflow in a2 - b2
    return ((a2 >= 0) == (b2 >= 0)) &&
            Math.abs(a2 - b2) <= representationDifference;
}

prints

1.1E-20 and 1.3E-20, similar= false
1.1E20 and 1.100000000000001E20, similar= true

Answer 2

public static boolean almostEqual(double a, double b, double eps){
    return Math.abs(a-b)<eps;
}

Where eps is measure of equality.

Answer 3

Approximate equality is defined in terms of the absolute difference: if an absolute difference does not exceed a certain, presumably small, number, then you can say that the values you are comparing are "close enough".

double diff = Math.abs(actual - expected);
if (diff < 1E-7) {
    // Numbers are close enough
}

You must be very careful to not confuse "close enough" end "equals", because the two are fundamentally different: equality is transitive (i.e. a==b and b==c together imply that a==c), while "close enough" is not transitive.

Answer 4

You must first decide what "almost the same" means. For example, there's a method in java.lang.Math called ulp() which, given a double, returns the distance between that double and the next; i.e., the smallest possible difference between that number and any other. You might simply compare the difference between the two doubles and the result of calling that method.

On the other hand, maybe you want two numbers to just be within 1% of eachother. In that case, do the same computation, but use the first number multiplied by 0.01 instead of ulp() as the largest acceptable distance.

Answer 5

You could use Guava and DoubleMath#fuzzyEquals method (since version 13.0):

public static boolean fuzzyEquals(double a, double b, double tolerance)

Returns true if a and b are within tolerance of each other. Technically speaking, this is equivalent to Math.abs(a - b) <= tolerance || Double.valueOf(a).equals(Double.valueOf(b)).

Notable special cases include:

Link to docs: https://google.github.io/guava/releases/17.0/api/docs/com/google/common/math/DoubleMath.html

Answer 6

What does it mean for two doubles to be "approximately equal?" It means that the doubles are within some tolerance of each other. The size of that tolerance, and whether that tolerance is expressed as an absolute number or as a percentage of the two doubles, depends on your application.

For instance, two photos displayed on a photo viewer have approximately the same width in inches if they take up the same number of pixels on the screen, so your tolerance will be an absolute number calculated based on pixel size for your screen. On the other hand, two financial firms' profits are probably "approximately equal" if they are within 0.1% of each other. These are just hypothetical examples, but the point is that it depends on your application.

Now for some implementation. Let's say your application calls for an absolute tolerance. Then you can use

private static final double TOLERANCE = 0.00001;

public static boolean approxEqual(final double d1, final double d2) {
    return Math.abs(d1 - d2) < TOLERANCE;
}

to compare two doubles, and use

approxEqual(d1, d2) && approxEqual(d1, d3) && approxEqual(d1, d4) && approxEqual(d1, d5)

to compare five doubles.