Files
pplus-library/modules/numerics/kahan-sum/routine.json
T
alexmandClaude df12b2a008 Start pplus-library: 16 verified routines in six modules
Routines written once in pseudo and checked at every level, so programs
can use them by name instead of rewriting and re-verifying them:

- stats: Mean, Variance (two-pass)
- numerics: KahanSum, LinearInterp, Trapezoid
- align: LinearGapPenalty, AffineGapPenalty, MatchScore,
  NeedlemanWunschScore
- md: LennardJones, LennardJonesForce, MinimumImage, PairDistance
- seq: GCContent, HammingDistance
- special: ErfApprox (Abramowitz and Stegun 7.1.26)

Each has contracts, cases, reference implementations in Python, Mojo
and (for the numerical ones) C at more than one level, and nine have a
Dafny proof over exact numbers. status.json and STATUS.md are written by
pplus-core's library-check: every check passes or is open/unverified as
listed there. The Mojo references still need the pinned compiler's
answers (library-verify-mojo, network).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FgvbRdoEheAHqsRmr2WuKZ
2026-10-05 11:58:22 +00:00

58 lines
1.5 KiB
JSON

{
"name": "KahanSum",
"summary": "The sum of a list of FLOAT values by compensated summation, whose error does not grow with the length of the list.",
"tags": [
"summation",
"rounding",
"precision"
],
"sources": [
{
"title": "Kahan, Further remarks on reducing truncation errors, CACM 8(1), 1965",
"url": "https://doi.org/10.1145/363707.363723"
},
{
"title": "Higham, Accuracy and Stability of Numerical Algorithms, section 4.3"
}
],
"cases": [
{
"call": "KahanSum([1.0, 2.0, 3.0])"
},
{
"call": "KahanSum([1.0, 1.0e-16, 1.0e-16, 1.0e-16, 1.0e-16, 1.0e-16, 1.0e-16, 1.0e-16, 1.0e-16, 1.0e-16, 1.0e-16])"
},
{
"call": "KahanSum([0.1, 0.1, 0.1, 0.1, 0.1, 0.1, 0.1, 0.1, 0.1, 0.1])"
},
{
"call": "KahanSum([])"
}
],
"references": [
{
"lang": "python",
"level": "plain",
"file": "ref/python/plain.py",
"note": "The same compensated loop."
},
{
"lang": "c",
"level": "plain",
"file": "ref/c/plain.c",
"note": "The same loop in C. It survives -O2; it would not survive -ffast-math, which lets the compiler treat (t - total) - y as 0."
},
{
"lang": "mojo",
"level": "plain",
"file": "ref/mojo/plain.mojo",
"note": "The same loop in Mojo."
}
],
"proof": {
"lang": "dafny",
"file": "proof/kahan_sum.dfy",
"note": "Over real numbers the compensation C stays 0 and the result is exactly the sum: what Kahan summation corrects is rounding, and nothing else."
}
}