multipleOf precision

[pashoo2]

For example, the result of ‌‌53.198098 % 0.0000001 is ‌4.068961201910684e-15. And this causes an errors with the "multipleOf" keyword.
Please, use something another instead of the modulo ("%") operator for validation of this keyword.

[epoberezkin]

It is also the case in v8 that 53.198098 / 0.0000001 = 531980980.00000006, although the precise value is 531980980.

There is no efficient way to improve integer arithmetics. Also I am not sure what is the practical reason to use "multipleOf": 0.0000001. If the number of decimal digits is really important maybe you can use a string with pattern instead...

Also you probably could define modulo precision for multipleOf as an option. It can be an option multipleOfPrecision: N (N is a positive integer). And in case it is defined, it will check that the modulo result is within -1e-N ... 1e-N range (instead of being exactly 0).

If that's what you want maybe you could make a PR with such option.

[blakeembrey]

@epoberezkin It might be worth looking at Number.EPSILON. That way the offset is factored into the calculation from the beginning (not sure if there's a drawback with using this, except maybe false positives on really tiny cases). This is only a slightly related comment in case you weren't aware of it.

[epoberezkin]

@blakeembrey thanks, it's interesting. Never heard of it :) The disadvantages are that it's ES6, so won't work in node 0.10 (we still use it...) and in some supported browsers, and that the above would fail anyway: 53.198098 % 0.0000001 is greater than Number.EPSILON.

[blakeembrey]

@epoberezkin Right, you'd use it to check the difference so you probably can't use it here (maybe, bad at math, sorry!). Also, it's just a number so I'm sure it's fine to polyfill. E.g. Math.pow(2, -52) === Number.EPSILON.

Edit: Thought, probably wrong, but you could do 53.1980435435 % 0.0000001 < Math.pow(10, -7) where -7 is the number of decimal places of accuracy.

[epoberezkin]

yes, 1e-7 will work too (it's compiled so can be x > -1e-{{=precision}} && x < 1e-{{=precision}} - modulo is negative for negative numbers)

[aaronshaf]

Not sure if this is related, but multipleOf 0.01 fails to validate 4.18, etc.

[epoberezkin]

Yes, same story here. 4.18 / 0.01 = 417.99999999999994. 4.18 % 0.01 is 0.01which is really bad... (multipleOf doesn't use modulo though, it compares the result of the division with parseInt() of this result).

The option I suggested would resolve this issue and probably it should be set to 12 by default...

[Palisand]

For the following property:

    "latitude": {
      "type": "number",
      "maximum": 85,
      "minimum": -85,
      "multipleOf": 1e-8
    }

Validation will fail for the value 40.7593657 even when multipleOfPrecision is set to 8.

[epoberezkin]

It passes with 6. multipleOfPrecision is explained in docs, see the docs on the information about this option.

[Palisand]

I read the option description in the docs before posting that comment. It makes no mention of 6 being the upper limit and states this option can be set "to some positive integer N". Did you get 6 from Math.abs(Math.round(division) - division) < 1e-N?

[epoberezkin]

You can't really compute this N, it just determines how close the result of division should be to a nearest integer number. The number of correct fractional digits depends on the number of digits in the integer part. So the correct N would depend on the range of the division results. With the divider 1e-8 you can't have too many fractional digits as the result is quite large...

There can be a better algorithm than Math.abs(Math.round(division) - division) < 1e-N, e.g. see #652 submitted today too.