BNL Benchmark Investigation

Lifecycle

archived

Owner

Ownership decision required

Review

No review record

Freshness

frozen-evidence

Evidence

captured 2026-05-25 · 4fdbf430b29edd24b9e1292c07ab24a151427dc7

Verified scopes

No verified code scopes recorded

Status: Archived, non-reproducible performance investigation

This record preserves measurements captured on 2026-05-25 at evidence revision 4fdbf430b29edd24b9e1292c07ab24a151427dc7. The BNL card, DOT graph, and ratio script remain, but the generated examples/cli/BNL/states/ directory, named workspaces, and result logs do not. The command blocks named below are also absent from the current card. Treat the values as historical evidence; reproducing them requires reconstructing the evidence-era setup.

This note records the BNL setup, generated states, and runtime/integration scans used for the 2 -> 5 amplitude graph in examples/cli/BNL/BNL.dot. All heavy generation and integration runs were executed behind a 30 GiB RSS guard. The final selected setup is CFF, SingleParametric, spherical coordinates with linear mapping, summed graphs, summed orientations, no LMB multichanneling, one Euler-angle stability rotation, and Double+Quad stability levels.

Kinematics and helicities

The final scan uses the same massless 2 -> 5 phase-space point and helicity assignment for all four masses; only MT and ymt are changed.

e_cm = 1000.0
helicities = [-1, 1, -1, 1, 1, -1, 1]
momenta = [
  [500.0, 0.0, 0.0, 500.0],
  [500.0, 0.0, 0.0, -500.0],
  [163.5141597587927, -155.7860143642699, -12.32488779805292, 48.12167194606676],
  [146.5267245560251, -18.94527369806406, 33.44069886802352, -141.3961713514227],
  [162.3143263141399, 4.025183527597962, -151.4903823484999, 58.14122874765845],
  [296.8259181161096, 175.6763342487531, 137.9119270904214, -195.5089553360614],
  [230.8188712549325, -4.970229714017088, -7.53735581189216, 230.6422259937584],
]

The kinematic point was selected from the random 2 -> 5 all-massless samples tested during the scan. The selected helicities flip the two initial leptons and the outgoing muon pair relative to the original trial point; this avoided a suppressed below-threshold central value and gave substantially smaller relative errors.

Threshold localization knobs

Two independent threshold-localization widths matter:

• Integrated threshold CT: subtraction.integrated_ct_settings.range.h_function_settings.sigma.
• Local threshold CT: subtraction.local_ct_settings.uv_localisation.gaussian_width.

No new setting was needed for the local width. The local counterterm uses evaluate_uv_damper, with gaussian_width as the width and sliver_width as the hard support gate. Event-level inspect checks confirmed that changing gaussian_width changes individual threshold_counterterm:* additional weights while leaving the original contribution unchanged.

The selected production values are:

subtraction.integrated_ct_settings.range.h_function_settings.sigma = 1.0
subtraction.local_ct_settings.uv_localisation.gaussian_width = 1.0
subtraction.local_ct_settings.uv_localisation.sliver_width = 10.0

Important scan observations:

• sigma = 0.25 was unstable for the selected threshold point.
• sigma = 4.0 sometimes reduced variance but changed the finite-statistics central value strongly in small samples, so it was not used for the final four-mass comparison.
• gaussian_width in the range 0.5..4.0 was viable; 1.0 was retained as a stable default and is the value used in the final integrations.
• cpe_iterations = 10000 stayed within generation budget but was slower at runtime than the optimized cpe_iterations = 5 state.
• A summed_function_map = true state was generated, but the corresponding SummedFunctionMap runtime path failed with an invalid helicity-type error in this branch, so it was not used for the final numbers.

Generated states

At capture time, the clean final states were kept under examples/cli/BNL/states/. That generated directory is not part of the retained repository evidence.

The evidence-era state names and generation invocations were:

• N: final_no_threshold_reference; run generate_bnl_no_threshold_reference -D mt=510.0 -D ymt=510.0
• N-opt: final_no_threshold_reference_optimized; run generate_bnl_no_threshold_reference_optimized -D mt=510.0 -D ymt=510.0
• T: final_threshold_reference; run generate_bnl_threshold_reference -D mt=400.0 -D ymt=400.0
• T-opt: final_threshold_reference_optimized; run generate_bnl_threshold_reference_optimized -D mt=400.0 -D ymt=400.0

Recorded generation command:

./target/release/gammaloop --clean-state --no-save-state \
  -s ./examples/cli/BNL/states/final_threshold_reference_optimized \
  ./examples/cli/BNL/BNL.toml \
  run generate_bnl_threshold_reference_optimized -D mt=400.0 -D ymt=400.0

The generation command blocks call save state -o explicitly. --no-save-state only prevents an extra save attempt at process exit.

Inspect benchmark timings

All timings below used inspect --bench 5 --n_batches 10 --minimal-integrand. The capture-time warmup-based auto-sizing undershot the requested wall time on this point, but the reported uncertainty is still the standard error over 10 batches.

Recorded benchmark commands:

./target/release/gammaloop --read-only-state \
  -s ./examples/cli/BNL/states/final_no_threshold_reference_optimized \
  ./examples/cli/BNL/BNL.toml \
  run bench_bnl_reference \
  -D mt=510.0 -D ymt=510.0 \
  -D disable_threshold_subtraction=true \
  -D output_file=final_no_threshold_reference_optimized_bench.json

./target/release/gammaloop --read-only-state \
  -s ./examples/cli/BNL/states/final_threshold_reference_optimized \
  ./examples/cli/BNL/BNL.toml \
  run bench_bnl_reference \
  -D mt=400.0 -D ymt=400.0 \
  -D disable_threshold_subtraction=false \
  -D integrated_h_sigma=1.0 \
  -D local_gaussian_width=1.0 \
  -D output_file=final_threshold_reference_optimized_bench.json

Final four-mass integrations

The final integration scan uses only the optimized setup. R1 and R2 are above threshold and keep threshold subtraction active. R3 and R4 are below threshold and disable threshold subtraction.

Recorded integration commands:

./target/release/gammaloop --read-only-state \
  -s ./examples/cli/BNL/states/final_threshold_reference_optimized \
  ./examples/cli/BNL/BNL.toml \
  run integrate_bnl_reference \
  -D mt=400.0 -D ymt=400.0 \
  -D disable_threshold_subtraction=false \
  -D integrated_h_sigma=1.0 \
  -D local_gaussian_width=1.0 \
  -D workspace=final_bnl_R1_mt400_threshold_2M_5c \
  -D n_start=200000 -D n_max=2000000 -D seed=990401 -D n_cores=5

./target/release/gammaloop --read-only-state \
  -s ./examples/cli/BNL/states/final_no_threshold_reference_optimized \
  ./examples/cli/BNL/BNL.toml \
  run integrate_bnl_reference \
  -D mt=510.0 -D ymt=510.0 \
  -D disable_threshold_subtraction=true \
  -D workspace=final_bnl_R3_mt510_no_threshold_3M_5c \
  -D n_start=300000 -D n_max=3000000 -D seed=990403 -D n_cores=5

The other two final commands are the same with mt=450.0, seed=990402, workspace=final_bnl_R2_mt450_threshold_2M_5c, threshold subtraction enabled; and mt=550.0, seed=990404, workspace=final_bnl_R4_mt550_no_threshold_3M_5c, threshold subtraction disabled.

Ratios

At capture time, BNL.py read the four final integration_result.json files by default and computed the ratios with first-order propagated independent real/imaginary uncertainties. Those default workspaces are absent from this checkout. The recorded invocation was:

python3 ./examples/cli/BNL/BNL.py \
  --json-output ./examples/cli/BNL/outputs/BNL_final_ratios.json

Final ratios:

Extended precision and corroboration update

The old final workspaces could not be resumed after the later branch changes: the integration workspace manifests contain generated-integrand fingerprints that no longer match the current loaded generated states. The resume guard correctly rejected this with:

Workspace integrand fingerprints do not match the current generated integrands
for BNL@default.

For that reason, the high-statistics extension below was run as a fresh current-branch workspace, not by bypassing the fingerprint check.

The corresponding workspaces were current_bnl_R1_mt400_threshold_10M_5c, final_bnl_R2_mt450_threshold_2M_5c, final_bnl_R3_mt510_no_threshold_3M_5c, and final_bnl_R4_mt550_no_threshold_3M_5c. None is retained in the repository.

The best independent-error ratios available in this record are:

This improves the previous ratios but does not reach the requested 0.25% relative uncertainty for every real and imaginary component. The bottleneck is the small real part of R1. After 10M samples its uncertainty is still 4.76e-18; propagating independent errors gives a 3.74% relative uncertainty on Im(R2/R1). Reaching 0.25% by brute-force independent sampling would require roughly (3.74 / 0.25)^2 ~= 224 times more R1 statistics, i.e. about 2.2e9 R1 samples at the observed variance. At the measured runtime this is an order-100-hour computation on five cores, so the ratio target is not practical with the independent-result estimator used in this record alone. A correlated multi-slot ratio estimator, or a variance-reduced estimator specifically for the small R1 real component, is needed for that precision target.

Corroboration runs were also made with independent seeds and safe runtime variations. The threshold-subtracted checks keep the UV scale and threshold localization fixed and only vary the spherical mapping parameter, because changing m_uv or threshold-localization widths changes the low-statistics central values in this local-subtraction-only setup and is not a clean corroboration knob.

The missing workspace names were corrob_map_R1_mt400_threshold_b0p8_500k_5c, corrob_map_R2_mt450_threshold_b1p2_500k_5c, corrob_phys_R3_mt510_no_threshold_b0p8_500k_5c, and corrob_phys_R4_mt550_no_threshold_graph_orientation_monte_carlo_500k_5c.