Number.MAX_SAFE_INTEGER incorrect?

# Anders Rundgren (21 days ago)

If you write Number.MAX_SAFE_INTEGER into a browser console it will return 9007199254740991

I believe this is wrong, 9007199254740992 the largest correct safe integer using IEEE-754 double precision.

Using my own IEEE-754 debugger:

Input floating point: 9007199254740991 Hex value: 433fffffffffffff Binary value: 0 10000110011 1111111111111111111111111111111111111111111111111111

Input floating point: 9007199254740992 Hex value: 4340000000000000 Binary value: 0 10000110100 0000000000000000000000000000000000000000000000000000

Using an external IEEE-754 debugger: www.binaryconvert.com/result_double.html?decimal=057048048055049057057050053052055052048057057050

Anders

# Isiah Meadows (21 days ago)

Technically, "safe" in this context means "can you store this number and all numbers below it without loss of precision", and for every single one of these numbers, you can still add one to them. When you get up to 2^53, you can no longer add just 1 - you can only add 2 or more. This is based on the number of significand bits - 2^53 - 1 has all lower 53 bits set and its exponent set to 0 (ignoring bias). You can precisely store up to this number without having to adjust the exponent, no more. This is what is meant by "safe". (The terminology of "exactly representable" is a misnomer

  • the highest integer you can theoretically represent without loss of precision is (2^53-1) * (2^971).)

This SO answer should help explain the situation better: stackoverflow.com/a/1848762


Isiah Meadows me at isiahmeadows.com, www.isiahmeadows.com

# Anders Rundgren (21 days ago)

On 2018-05-06 18:40, Isiah Meadows wrote:

Technically, "safe" in this context means "can you store this number and all numbers below it without loss of precision", and for every single one of these numbers, you can still add one to them. When you get up to 2^53, you can no longer add just 1 - you can only add 2 or more. This is based on the number of significand bits - 2^53 - 1 has all lower 53 bits set and its exponent set to 0 (ignoring bias). You can precisely store up to this number without having to adjust the exponent, no more. This is what is meant by "safe". (The terminology of "exactly representable" is a misnomer - the highest integer you can theoretically represent without loss of precision is (2^53-1) * (2^971).)

Thanx Isiah, I (sort of) understand :-)

Anyway, this definition departs from similar definitions in other languages and AFAICT, existing JavaScript implementations have no problems using 2^53 as an integer.

var t = 9007199254740992 < undefined t-1 < 9007199254740991 t < 9007199254740992

For JavaScript this may be of limited importance but for (I-)JSON canonicalization [1] it is vital that all parties do the same interpretation. That is, for I-JSON the 2^53-1 limitation is simply put wrong.

Anders

1] cyberphone/json-canonicalization#json

# Logan Smyth (21 days ago)

To clarify, the "you can still add one to them" is probably the most important part in that definition, given your question.

Anyway, this definition departs from similar definitions in other

languages and AFAICT, existing JavaScript implementations have no problems using 2^53 as an integer.

2^53 can be represented just fine, but 2**53 + 1 === 2**53 is true, so adding one to it does not actually increment the value, meaning it does not satisfy the definition of SAFE in this context.

I think the best source of truth is likely the spec: www.ecma-international.org/ecma-262/8.0/#sec-number.max_safe_integer which states

The value of Number.MAX_SAFE_INTEGER is the largest integer n such that n

and n + 1 are both exactly representable as a Number value.

# Mark Miller (21 days ago)

motivation:

if (isSafeInteger(a) && isSafeInteger(b)) { const c = a + b; if (isSafeInteger(b)) { // c is actually the sum of a and b } }

# Mark Miller (21 days ago)

Oops. Should be:

if (isSafeInteger(a) && isSafeInteger(b)) { const c = a + b; if (isSafeInteger(c)) { // c is actually the sum of a and b } }

# Anders Rundgren (21 days ago)

On 2018-05-06 19:57, Logan Smyth wrote: <snip>

I think the best source of truth is likely the spec: www.ecma-international.org/ecma-262/8.0/#sec-number.max_safe_integer which states

The value of Number.MAX_SAFE_INTEGER is the largest integer n such that n and n + 1 are both exactly representable as a Number value.

Right, this is essentially what I'm claiming; Number.MAX_SAFE_INTEGER + 1 is a valid (exact) integer which means that tools.ietf.org/html/rfc7493#section-2.2 is incorrect.

Anders

# Mark Miller (21 days ago)

you are correct. The rfc as stated is incorrect. The EcmaScript spec is correct.

253 is indeed exactly representable, which is what the rfc is about. But 253 is not safe, which what the ecmascript spec is about.

# Mark Miller (21 days ago)

I take that back. The rfc says:

An I-JSON sender cannot expect a receiver to treat an integer whose absolute value is greater than 9007199254740991 (i.e., that is outside the range [-(253)+1, (253)-1]) as an exact value.

For the natural interpretation of "treat" as in "operate on or with" I'd say the rfc is correct. But the language is ambiguous and should be clarified.

# Anders Rundgren (21 days ago)

On 2018-05-06 21:37, Mark Miller wrote:

I take that back. The rfc says:

An I-JSON sender cannot expect a receiver to treat an integer whose
absolute value is greater than 9007199254740991 (i.e., that is
outside the range [-(2**53)+1, (2**53)-1]) as an exact value.

For the natural interpretation of "treat" as in "operate on or with" I'd say the rfc is correct. But the language is ambiguous and should be clarified.

Well, the author questioned that IEEE-754 implementations actually deal with this edge case at all.

However, they do which I found out by testing with 9007199254740992 in order to break my canonicalizer implementations which to my surprise all continued to work.

Then, when looking at the binary, it became clear that Number.MAX_SAFE_INTEGER is not comparable to other languages' MAX* definitions. docs.oracle.com/javase/8/docs/api/constant-values.html#java.lang.Integer.MAX_VALUE

Anyway, thanx for the insights in this matter!

Anders