Project

General

Profile

Feature #15811

Propsing new method for comparing equality of 2 (float) numbers relatively

Added by yennguyenh (yen nguyen) about 2 months ago. Updated about 1 month ago.

Status:
Open
Priority:
Normal
Assignee:
-
Target version:
-
[ruby-core:92475]

Description

Background

Equal comparison method between 2 float numbers returns unexpected results sometimes. Therefore, a relative comparison method is needed!

Proposal

A relative equal comparison method has been written based on a Python project! This method gives the approximation for the equal comparison based on two values: realative tolerance and absolute tolerance. Near zero value will also be considered carefully!

Implementation

The function for that would be called close?
close?(a, b, rel_tol, abs_tol)

a and b: are the two values to be tested to relative closeness

rel_tol: is the relative tolerance -- it is the amount of error allowed, relative to the larger absolute value of a or b. For example, to set a tolerance of 5%, pass tol=0.05. The default tolerance is 1E-9, which assures that the two values are the same within about 9 decimal digits. rel_tol must be greater than 0.0

abs_tol: is a minimum absolute tolerance level -- useful for comparisons near zero.

Evaluation of your implementation

By default, relative tolerance is 1E-9 which is relatively precise enough to compare two float numbers. However it can also be adjusted in case higher accuracy is requested. The absolute tolerance is by default 0.0 and need to be set in case of near-zero numbers.

Discussion

There are some test cases available for the method which has approved the accuracy of the method. BigNumbers and integers are also tested. However, more test cases are still needed to assure even better the accuracy of the method.

Gist

Relative equal comparison

https://gist.github.com/yennguyenh/63d5e7a11f354f796b43ada037c4b2c5

Test cases

https://gist.github.com/yennguyenh/2e81dc72b310cb9d886a82faf3d536ef


Related issues

Related to Ruby trunk - Feature #10425: A predicate method to tell if a number is near anotherOpen10/26/2014Actions

History

Updated by yennguyenh (yen nguyen) about 2 months ago

I don't know how to add the new method correctly. Please have a look at those gist and tell me what to do! Thank you :)

Updated by shevegen (Robert A. Heiler) about 2 months ago

Not sure if the name .close? is a good name to indicate a relative comparison. Note that
I have no real pro or con opinion, just pointing out that the name may not be ideal. I
don't have a good alternative proposal either; perhaps something a bit longer, two words?

Perhaps you could add a more specific example - might be useful. Will this method reside
on Math , for example? E. g.:

Math.close?(a, b, rel_tol, abs_tol)

Again, not having any preference myself here, just asking to see more details added to
the suggestion. The ruby team is open to adding/discussing changes/modifications based
on use cases; in my opinion, the more details can be added (when they are important),
the better.

To make it easier for others, I copy/pasted a part of your gist sample (excluding the
documents), just to make it simpler for people to read the code here; you can indent
via 4 ' ' and then the bugtracker here will correctly highlight the ruby code:

def self.close?(a, b, rel_tol: RELATIVE_TOLERANCE, abs_tol: ABSOLUTE_TOLERANCE)
  raise ArgumentError.new('Arguments must be numeric') unless (a.is_a?(Numeric) && b.is_a?(Numeric))
  raise ArgumentError.new('Error tolerance must positive') if (rel_tol < 0.0 || abs_tol < 0.0)

  # short-cut exact equality
  return true if a == b

  # check if any attribute is Infinite
  return false if a.infinite? || b.infinite?

  # weak comparition - the tolerance is scaled by the larger of 2 values
  abs_diff = (a - b).abs
  ((abs_diff <= (rel_tol * b).abs) ||
    (abs_diff <= (rel_tol * a).abs) ||
    (abs_diff <= abs_tol))
end

Updated by wishdev (John Higgins) about 2 months ago

The tests for this are incorrect and show why this does not work.

From the test gist

context 'Numbers between 1 and 0' do
  let (:absolute_tolerance) { 1.0E-14 }
  it 'returns true for same positive numbers' do
    expect(Math.close?(1000001.0E-15, 1000002.0E-15, abs_tol: absolute_tolerance)).to be_truthy
    expect(Math.close?(1000002.0E-15, 1000001.0E-15, abs_tol: absolute_tolerance)).to be_truthy
  end
  it 'returns true for same negative numbers' do
    expect(Math.close?(-1000001.0E-15, -1000002.0E-15, abs_tol: absolute_tolerance)).to be_truthy
    expect(Math.close?(-1000002.0E-15, -1000001.0E-15, abs_tol: absolute_tolerance)).to be_truthy
  end
  it 'returns false for different positive numbers' do
    expect(Math.close?(1000010.0E-15, 1000020.0E-15, abs_tol: absolute_tolerance)).to be_falsey
    expect(Math.close?(1000020.0E-15, 1000010.0E-15, abs_tol: absolute_tolerance)).to be_falsey
  end
  it 'returns false for different negative numbers' do
    expect(Math.close?(-1000010.0E-15, -1000020.0E-15, abs_tol: absolute_tolerance)).to be_falsey
    expect(Math.close?(-1000020.0E-15, -1000010.0E-15, abs_tol: absolute_tolerance)).to be_falsey
  end
end

10E-15 == 1E-14 therefore since the absolute tolerance is equal to the difference of the bottom two "return false" specs - they must be true - they are not true because subtracting those floats ends up with garbage.

For example

1000020.0E-15 - 1000010.0E-15

equals

1.0000000000085785e-14

Which places it outside of 1E14 but common sense (and looking at the numbers in front of us) obviously the correct answer is 1E14.

Floating numbers cannot be acted upon and then the result used to prove something.

This does not provide what it claims to provide - it is not possible to provide what you wish to provide here when dealing with floats.

Sorry

John

Updated by nobu (Nobuyoshi Nakada) about 2 months ago

  • Description updated (diff)

I think it should be under Math or Float, and a independent gem could be a good first step.

Updated by duerst (Martin Dürst) about 2 months ago

Ruby is an object-oriented language. So I think this should be something like:

a.close_to?(b, abs_tolerance: t)

or so, not a function with two main numbers.

#8

Updated by mame (Yusuke Endoh) about 2 months ago

  • Related to Feature #10425: A predicate method to tell if a number is near another added

Updated by nobu (Nobuyoshi Nakada) about 2 months ago

duerst (Martin Dürst) wrote:

Ruby is an object-oriented language. So I think this should be something like:

a.close_to?(b, abs_tolerance: t)

or so, not a function with two main numbers.

If it is an instance method, the relative tolerance feels relative to the absolute value of the receiver, not the larger one.

Updated by yennguyenh (yen nguyen) about 1 month ago

nobu (Nobuyoshi Nakada) wrote:

I think it should be under Math or Float, and a independent gem could be a good first step.

I have updated the first gist. It is under Math! I just forgot putting it in gist . Thank you for reminding anyway :D

#11

Updated by yennguyenh (yen nguyen) about 1 month ago

wishdev (John Higgins) wrote:

The tests for this are incorrect and show why this does not work.

From the test gist

context 'Numbers between 1 and 0' do
  let (:absolute_tolerance) { 1.0E-14 }
  it 'returns true for same positive numbers' do
    expect(Math.close?(1000001.0E-15, 1000002.0E-15, abs_tol: absolute_tolerance)).to be_truthy
    expect(Math.close?(1000002.0E-15, 1000001.0E-15, abs_tol: absolute_tolerance)).to be_truthy
  end
  it 'returns true for same negative numbers' do
    expect(Math.close?(-1000001.0E-15, -1000002.0E-15, abs_tol: absolute_tolerance)).to be_truthy
    expect(Math.close?(-1000002.0E-15, -1000001.0E-15, abs_tol: absolute_tolerance)).to be_truthy
  end
  it 'returns false for different positive numbers' do
    expect(Math.close?(1000010.0E-15, 1000020.0E-15, abs_tol: absolute_tolerance)).to be_falsey
    expect(Math.close?(1000020.0E-15, 1000010.0E-15, abs_tol: absolute_tolerance)).to be_falsey
  end
  it 'returns false for different negative numbers' do
    expect(Math.close?(-1000010.0E-15, -1000020.0E-15, abs_tol: absolute_tolerance)).to be_falsey
    expect(Math.close?(-1000020.0E-15, -1000010.0E-15, abs_tol: absolute_tolerance)).to be_falsey
  end
end

10E-15 == 1E-14 therefore since the absolute tolerance is equal to the difference of the bottom two "return false" specs - they must be true - they are not true because subtracting those floats ends up with garbage.

For example

1000020.0E-15 - 1000010.0E-15

equals

1.0000000000085785e-14

Which places it outside of 1E14 but common sense (and looking at the numbers in front of us) obviously the correct answer is 1E14.

Floating numbers cannot be acted upon and then the result used to prove something.

This does not provide what it claims to provide - it is not possible to provide what you wish to provide here when dealing with floats.

Sorry

John

Sorry but I do not really understand what you meant. What I get so far is that you mean the difference of that pair of number (1000020.0E-15 - 1000010.0E-15) results not as expected, 1.0000000000085785e-14 instead of 1.0e-14. I have taken a look on that and realize one mistake on the algorithm. The absolute tolerance is set to check the accuracy to a certain decimal place and so at that place the difference should be less than 1 which is 0. Therefore the equal case should not be considered as the case for equal numbers. It should be fixed like below ( I have also updated the code!)

current method:

abs_diff = (a - b).abs
    ((abs_diff <= (rel_tol * b).abs) ||
      (abs_diff <= (rel_tol * a).abs) ||
      (abs_diff <= abs_tol))

fixed method:

abs_diff = (a - b).abs
    ((abs_diff <= (rel_tol * b).abs) ||
      (abs_diff <= (rel_tol * a).abs) ||
      (abs_diff < abs_tol))

For ex:

absolute tolerance: 1e-2
a: 0.01
b: 0.02
(a-b).abs: 0.001 == 1e-2
At the second decimal place, there is the difference of '1' which should return false for the equal comparison, so it return false in case absolute tolerance == (a-b).abs

If it is not what you meant, please explain me more! Anyway thank you for the feedback, that I could find out that mistake!

Also available in: Atom PDF