thinking about floating point and accuracy tradeoffs in numerical computation.
binary floating point is steppy, with gaps between representable numbers occasionally doubling.
wondering if log-domain would work better/worse in different applications? instead of mantissa and exponent, just have a fractional base and put everything in the exponent.
should make possible smoother tradeoffs between precision and range.
probably not feasible to implement in hardware though. might try some experiments some time with low precision using lookup tables.
claude
npub14z…y5gys
2024-10-30 20:34:50 UTC
Author Public Key
npub14zz93nwc6c9267yl4ksgydkzy2087yzue8gpmww2nr5d8q6waqlq2y5gysSeen on
wss://relay.momostr.pinkPublished at
2024-10-30 20:34:50 UTCEvent JSON
{
"id": "bf59ccdbd18c83fa0d104c66dced826fc24a402b151b5aeb55dd5aecf30a903e",
"pubkey": "a88458cdd8d60aad789fada08236c2229e7f105cc9d01db9ca98e8d3834ee83e",
"created_at": 1730320490,
"kind": 1,
"tags": [
[
"proxy",
"https://post.lurk.org/@mathr/113398283665361737",
"web"
],
[
"proxy",
"https://post.lurk.org/users/mathr/statuses/113398283665361737",
"activitypub"
],
[
"L",
"pink.momostr"
],
[
"l",
"pink.momostr.activitypub:https://post.lurk.org/users/mathr/statuses/113398283665361737",
"pink.momostr"
],
[
"-"
]
],
"content": "thinking about floating point and accuracy tradeoffs in numerical computation.\n\nbinary floating point is steppy, with gaps between representable numbers occasionally doubling.\n\nwondering if log-domain would work better/worse in different applications? instead of mantissa and exponent, just have a fractional base and put everything in the exponent.\n\nshould make possible smoother tradeoffs between precision and range.\n\nprobably not feasible to implement in hardware though. might try some experiments some time with low precision using lookup tables.",
"sig": "5f3d89414828f61e855ef25434f93d06024095c2be68ffc9c9863943ae9e104dc485d9b5e1b692d95bc96abb7b312ac26b8feb4beaabba8430e7695f6827f5ef"
}