Re: [Json] "best practices" Vs. Profile for i-json

[Stephen Dolan <[e-mail]>]

On Fri, Aug 1, 2014 at 10:55 PM, Larry Masinter <[e-mail]> wrote:
> "All numeric values in I-JSON MUST round-trip to IEEE floating point and back to the same string."

The exact requirement is more subtle than this.

The number 1.1 cannot be represented exactly as an IEEE754 double. The
closest number which can is:

A = 1.100000000000000088817841970012523233890533447265625

The requirements that most languages make of their double conversion
functions (at least, those that bother to specify such things) are:

(1) parsing a string returns the IEEE754 double which is closer to the
decimal number than any other IEEE754 double (with rules for breaking
ties, and rules for strings representing numbers outside [-DBL_MAX,
DBL_MAX])

(2) printing a double returns some string whose decimal value is
closer to the provided IEEE754 double than to any other IEEE754 double
(with similar exceptions for ties, nan, inf, etc.)

Some languages (e.g. Java) go further and require that printing return
the shortest string compatible with (2). Without this condition, there
are many different strings that validly correspond to A above: "1.1"
(the shortest), "1.10", "1.10000000000000009". In general,
print(parse(str)) need not be equal to str, but
parse(print(parse(str))) must be equal to parse(str).

The point of this is that being a valid string representation of an
IEEE754 number has nothing to do with round-tripping to the same
string, and is not a property of the string itself. Instead, the
condition that specifies "use doubles with json" has to do with when
JSON values are considered equivalent.

Two JSON documents differing only in whitespace outside strings are
(MUST be? SHOULD be?) considered equivalent by any implementation. If
you want to specify "use doubles" you should state something similar,
like:

"An implementation [MUST NOT/SHOULD NOT] distinguish between two
numeric literals whose decimal values round to the same IEEE754
double-precision value (using standard round-to-nearest, ties-to-even
rounding)."

Stephen