My a few additional notes to random floating-point number API.

[Zhu Zihao]

Hi, everybody.

Since we have not yet finalized the minutes of 03-23 WG2 meeting. I'd
like add one more point about random floating-point number API.

In the meeting, John Cowan demonstrated a part about random
floating-point number of CLtL
(<https://www.cs.cmu.edu/Groups/AI/html/cltl/clm/node133.html>)

CLtL noted that, The accepted pratice of generating a random
floating-point in range [A ,B) is first generate a qualified random
floating-point in range [0, 1), then use affine transformation X * (B -
A) + A to scale it.

However, Frédéric Goualard in his paper "Drawing random floating-point
numbers from an interval" (<https://hal.science/hal-03282794v2>) points
out that: due to the specific properties of IEEE-754, the affine
transformation above will map serveral floats in [0, 1), to the same
float in [A, B) in some cases, thus break the uniformity.

See the Section 2 of his paper for more formal and detailed explanation.

Therefore, I believe that a potential random floating-point API should
allow users to specify custom intervals, rather than being limited to
[0, 1).

P.S.: When compiling the minutes, please feel free to include the points
here as an addendum if necessary, Thanks :)
--
Retrieve my PGP public key:
执行下列命令以获取我的 PGP 公有密钥：

gpg --recv-keys B3EBC086AB0EBC0F45E0B4D433DB374BCEE4D9DC

Zihao / 閱卜錄