verification/data/claim5_source_excerpt.txt

Source: https://ar5iv.labs.arxiv.org/html/2605.31369
Retrieved: 2026-07-29T11:22:44Z
Full HTML SHA-256: 9e465ccada2014404d315c46108399b4ee49693d09d9866ff3dbd426045fd30a
Section 5.3:
"For f-divergences, JKO yields moderate yet consistent improvements."

The complete reported CIFAR-10 FID table:
Divergence,noJKO,tau=0.01,tau=1,tau=100
ckMMD,13.96,129.38,12.78,14.12
chi2,15.60,22.26,15.01,15.87
KL,15.78,26.94,15.47,17.28
KL-DV,16.68,25.19,16.56,16.88
MMD,49.39,28.36,15.76,21.43
Shannon,14.60,32.24,15.92,15.23
sMMD,59.79,135.27,26.79,55.59
Wasserstein-1,35.37,22.66,28.91,35.35

Appendix complementary table:
"The runs used in this table correspond to different random seeds than those
reported in Table 4, which explains the slight differences in FID values."
Divergence,noJKO,tau=0.01,tau=1,tau=100
ckMMD,13.16,241.52,13.42,14.41
chi2,17.93,23.16,17.39,17.63
KL,17.62,34.54,18.30,19.13
KL-DV,17.98,24.46,18.72,18.86
MMD,25.48,28.16,16.39,46.62
Shannon,16.84,31.53,17.26,16.87
sMMD,62.58,106.66,25.73,47.09
Wasserstein-1,36.77,22.75,33.34,41.57

Authors' code mapping at revision 6633b553a244634bd2c2e1142603aad1c1fbe55a:
README.md: "f-divergences (KL, chi², Jensen-Shannon)"
README.md divergence flag: "KL_DV, KL, MMD, Shannon, chi2, Wasserstein-1"
losses.py labels the softplus GAN loss branch as "Shannon" and states that
"Jensen-Shannon" is the corresponding symmetric divergence.