Skip to content

test(nuts): bound the divergence rate instead of requiring exactly zero #215

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
MaartenMarsman merged 1 commit into from
Oct 21, 2026
Merged

test(nuts): bound the divergence rate instead of requiring exactly zero #215

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
8 changes: 6 additions & 2 deletions tests/testthat/test-ggm-nuts.R
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -10,6 +10,10 @@
# 1. Posterior moment comparison (means, variances, KS, bivariate)
# 2. Edge selection accuracy (PIPs)
# 3. Gradient near PD boundary
#
# Divergences: NUTS trajectories differ across platforms, so the blocks bound
# the divergence rate (below 0.1% of post-warmup transitions) rather than
# requiring exactly zero.
# 4. NUTS diagnostics are well-behaved
#
# TODO: add a proper Geweke JDT or SBC test for one-transition correctness
Expand Down Expand Up @@ -189,7 +193,7 @@ test_that("NUTS posterior means agree on p=6 tridiagonal", {
expect_lt(max(diff), 0.05, label = "max abs diff in pairwise means < 0.05")

# NUTS should have clean diagnostics
expect_equal(fit_nuts$nuts_diag$summary$total_divergences, 0)
expect_lt(fit_nuts$nuts_diag$summary$divergence_rate, 1e-3)
})


Expand Down Expand Up @@ -373,7 +377,7 @@ test_that("NUTS diagnostics are clean for well-specified model", {
)

expect_true(!is.null(fit$nuts_diag))
expect_equal(fit$nuts_diag$summary$total_divergences, 0)
expect_lt(fit$nuts_diag$summary$divergence_rate, 1e-3)
expect_equal(fit$nuts_diag$summary$max_tree_depth_hits, 0)
expect_gt(fit$nuts_diag$summary$min_ebfmi, 0.3)
})
Expand Down
20 changes: 12 additions & 8 deletions tests/testthat/test-mixed-nuts.R
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -11,6 +11,10 @@
# E conditional PL, edge selection, interleaved variable order
# F conditional PL, no edge selection, interleaved variable order
#
# Divergences: NUTS trajectories differ across platforms, so the blocks bound
# the divergence rate (below 0.1% of post-warmup transitions) rather than
# requiring exactly zero.
#
# Weekly certification tier (T2, BGMS_RUN_CERTIFICATION): a full NUTS-vs-MH
# condition grid, and each condition fits two
# chains x 5000 iterations for both samplers.
Expand Down Expand Up @@ -176,7 +180,7 @@ test_that("M.2A: NUTS vs MH agree (conditional PL, no ES, grouped)", {
diff = abs(fit_nuts$posterior_mean_pairwise - fit_mh$posterior_mean_pairwise)
expect_lt(max(diff), 0.1, label = "M.2A max abs diff in associations < 0.1")

expect_equal(fit_nuts$nuts_diag$summary$total_divergences, 0)
expect_lt(fit_nuts$nuts_diag$summary$divergence_rate, 1e-3)
})


Expand Down Expand Up @@ -211,7 +215,7 @@ test_that("M.2B: NUTS vs MH agree (marginal PL, no ES, grouped)", {
compare_pairwise_samples(fit_nuts, fit_mh, "M.2B")
compare_main_samples(fit_nuts, fit_mh, "M.2B")

expect_equal(fit_nuts$nuts_diag$summary$total_divergences, 0)
expect_lt(fit_nuts$nuts_diag$summary$divergence_rate, 1e-3)
})


Expand Down Expand Up @@ -256,7 +260,7 @@ test_that("M.2C: NUTS vs MH agree (conditional PL, ES, grouped)", {
label = "M.2C PIP correlation > 0.95"
)

expect_equal(fit_nuts$nuts_diag$summary$total_divergences, 0)
expect_lt(fit_nuts$nuts_diag$summary$divergence_rate, 1e-3)
})


Expand Down Expand Up @@ -300,7 +304,7 @@ test_that("M.2D: NUTS vs MH agree (marginal PL, ES, grouped)", {
label = "M.2D PIP correlation > 0.95"
)

expect_equal(fit_nuts$nuts_diag$summary$total_divergences, 0)
expect_lt(fit_nuts$nuts_diag$summary$divergence_rate, 1e-3)
})


Expand Down Expand Up @@ -377,7 +381,7 @@ test_that("M.2E: NUTS vs MH agree (conditional PL, ES, interleaved)", {
colnames(extract_pairwise_interactions(fit_mh))
)

expect_equal(fit_nuts$nuts_diag$summary$total_divergences, 0)
expect_lt(fit_nuts$nuts_diag$summary$divergence_rate, 1e-3)
})


Expand Down Expand Up @@ -431,7 +435,7 @@ test_that("M.2F: NUTS vs MH agree (conditional PL, no ES, interleaved)", {
colnames(extract_pairwise_interactions(fit_mh))
)

expect_equal(fit_nuts$nuts_diag$summary$total_divergences, 0)
expect_lt(fit_nuts$nuts_diag$summary$divergence_rate, 1e-3)
})


Expand Down Expand Up @@ -591,7 +595,7 @@ test_that("M.2J: NUTS diagnostics are clean for mixed MRF", {
)

expect_true(!is.null(fit$nuts_diag))
expect_equal(fit$nuts_diag$summary$total_divergences, 0)
expect_lt(fit$nuts_diag$summary$divergence_rate, 1e-3)
expect_equal(fit$nuts_diag$summary$max_tree_depth_hits, 0)
expect_gt(fit$nuts_diag$summary$min_ebfmi, 0.3)
})
Expand Down Expand Up @@ -635,5 +639,5 @@ test_that("M.2T: NUTS vs MH agree under tilt (conditional PL, no ES, delta=1)",
label = "M.2T max abs diff in associations < 0.1 (delta=1)"
)

expect_equal(fit_nuts$nuts_diag$summary$total_divergences, 0)
expect_lt(fit_nuts$nuts_diag$summary$divergence_rate, 1e-3)
})
Loading