V4 — on real phylogenies, does MCEM reach the exact conditioned maximum?
| Status | registered 2026-10-31; read 2026-11-01 on all 700 fits |
| Series | the validation series (replaces E1, E42 (E45's dd cells are re-scored in E45)) |
| Package build | emphasis 182055b (1.0.0.9000), R 4.5.3 |
| Driver | sim/V4.R |
| Data | results/V4/jobs/, results/V4/build.txt |
Question
Where an exact likelihood exists (cr, and dd at full sampling), MCEM's estimate can be scored against the exact maximum. E42 did this on 75 of E40's trees with an unconditioned old build and a dd reference later repaired (c641dca); V4 refits all 175 trees on the current build, conditioned.
Design
| trees | E40's 175: 116 class-tree clades (\(\rho = 1\)), 59 Condamine families at their sampling fraction |
| fits | cr, dd; default rule and mc_error; conditioned; auto_bounds() box |
| exact | cr conditioned (DDD at \(\rho = 1\), Bernoulli closed form below); dd at \(\rho = 1\), validity checked on the truncation |
| jobs | 700 |
Pre-set rules
- On trees with a valid exact maximum: median \(\Delta\ell \ge -0.5\) and 90th percentile of \(|\Delta\ell| \le 2\), per model, rule and set (E45's C1).
Results
sim/V4-analysis.py (results/V4/rule.csv, paired.csv): 700 fits, 610 with a valid exact conditioned maximum (cr on every tree; dd only at complete sampling, so 14 of the 59 family trees).
The rule holds in all eight groups. \(\Delta\ell = \ell(\hat\theta_{\text{MCEM}}) - \ell(\hat\theta)\):
| model | stopping rule | tree set | trees with an exact maximum | median \(\Delta\ell\) | 90th percentile of \(\lvert\Delta\ell\rvert\) | worst | on the iteration cap |
|---|---|---|---|---|---|---|---|
cr |
mc_error |
family trees | 59 | -0.04 | 0.13 | -0.44 | 14% |
cr |
mc_error |
class-tree clades | 116 | -0.10 | 0.29 | -0.44 | 36% |
cr |
default | family trees | 59 | -0.07 | 0.30 | -0.72 | 0% |
cr |
default | class-tree clades | 116 | -0.12 | 0.34 | -1.17 | 0% |
dd |
mc_error |
family trees | 14 | -0.01 | 0.41 | -14.84 | 0% |
dd |
mc_error |
class-tree clades | 116 | -0.01 | 0.56 | -2.89 | 1% |
dd |
default | family trees | 14 | -0.16 | 0.53 | -14.85 | 7% |
dd |
default | class-tree clades | 116 | -0.06 | 0.65 | -4.09 | 6% |
- On the 305 trees fitted under both rules,
mc_errorlands closer to the exact maximum on 75% (median \(\Delta\ell\) -0.05 against -0.09 for the default rule). - One tree stands apart: Picidae (223 tips) under
dd, where both rules stop at the same point, 14.8 below the exact maximum (15 iterations at 200 draws and 93 at 2400). The extinction slope on \(N\) is on the lower edge of theauto_bounds()box in all three fits (\(\gamma_N = -0.013453\)), and along that edge the likelihood has two modes: MCEM, started from the box's default, stops at \(\lambda_0 = 0.20\), \(\mu_0 = 0.16\); the exact maximiser, with restarts, reaches \(\lambda_0 = 0.72\), \(\mu_0 = 0.79\). The next worst are Pteropodidae (-4.1), Pitheciidae (-2.9) and Tenrecidae (-2.9), all under the default rule. - A
crfit that ends on the iteration cap undermc_error(36% of the class-tree clades) is still at the exact maximum to within the table's spread.
Conclusion
On real phylogenies MCEM reaches the exact conditioned maximum for cr and dd; the mc_error rule gets closer than the default. Picidae under dd is the exception: with the extinction slope on the box's edge, MCEM from the default start reaches a low-turnover mode 14.8 below the exact maximum's high-turnover one.