Skip to content

Skip PSIS tail smoothing when a quarter of the tail ties at its minimum - #3451

Merged
WardBrian merged 4 commits into
developfrom
fix-psis-zero-excesses
Nov 6, 2026
Merged

WardBrian merged 4 commits into
developfrom
fix-psis-zero-excesses

Conversation

@avehtari

Copy link
Copy Markdown
Member

This PR was assisted by Claude. The reference code in R was written by human.

Summary

Zhang and Stephens (2009) method requires that first quartile is bigger than the smallest value. psis_smooth_tail() passes the excesses exp(lw_tail) - exp(cutoff) to gpdfit(). When a quarter of them are 0, the first quartile x_1st_qt is 0, and the grid theta divides by it and becomes -Inf. The product -theta * x is then NaN for the zero excesses, so k and sigma are NaN. The test !std::isinf(k) lets NaN through, the smoothed tail is NaN, and (llr_weights.array() < 0.0).select(llr_weights, 0.0) maps NaN to 0. Each tail draw gets the weight exp(0) = 1. The k > 0.7 warning does not fire for NaN.

Zero excesses occur in these cases:

  • tail ratios tie with the cutoff, for example duplicate draws;
  • the cutoff and some tail ratios are more than about 745 below the largest ratio, so exp() underflows to 0 for them;
  • many draws have a log density of -Inf, which single.hpp sets when the log density throws a std::domain_error.

Example (100 draws, tail length 20): 70 log ratios from -20 to -26.9, 15 at -5, and 15 from -3.5 to 0. The cutoff and 5 tail values are -5. develop returns 1 for all 20 tail draws; the correct unsmoothed weights are exp(lr), for example exp(-5) = 0.0067 for the 5 tied draws.

When the tied minimum is positive (a quarter of the tail ties above the cutoff), theta is finite, and gpdfit() fits a sample that is far from a generalized Pareto distribution. posterior::gpdfit() returns NA for both cases, with the same check on the first quartile.

Intended Effect

  • gpdfit() returns NaN for sigma and k when the first quartile of x is not larger than its smallest value.
  • psis_smooth_tail() smooths only for a finite k. Otherwise it returns the tail unchanged.
  • psis_weights() logs a warning when k is not finite and returns the weights without smoothing, as it already does for a constant tail.

This behavior matches the behavior of corresponding function in posterior R package.

How to Verify

./runTests.py src/test/unit/services/pathfinder. The three new tests in psis_test.cpp are:

  • gpdfit_tied_first_quartile: 4 of 20 values at the minimum give a finite fit; 5 of 20 at a minimum of 0, and 5 of 20 at a minimum of 0.001, give NaN.
  • psis_weights_tied_cutoff: the example above. The weights must equal exp(lr), and the warning must be logged.
  • psis_weights_underflow_in_tail: 85 log ratios from -1000 to -1084 and 15 from -3.5 to 0, so 5 tail values and the cutoff underflow. The same checks.

With the develop psis.hpp and the new tests, the 3 new tests fail (42 expectations) and the 5 existing tests pass. With this PR all 8 pass. eight_schools_test and normal_glm_test also pass.

Tested with develop 849cec8da and Stan Math e4745c87ff, g++ 14.3.0, Linux x86_64.

Side Effects

The results do not change when the first quartile of the excesses is larger than their minimum. The existing PSIS tests and the Pathfinder service tests pass without changes. A tail with a quarter of its values tied at a positive minimum is no longer smoothed.

Documentation

  • The doc comments of gpdfit() and psis_smooth_tail() describe the new return values.
  • The second commit fixes the capitalization of Pareto in docs, warnings and comments (no code changes)

Copyright and Licensing

Please list the copyright holder for the work you are submitting (this will be you or your assignee, such as a university or company):

Aalto University

By submitting this pull request, the copyright holder is agreeing to license the submitted work under the following licenses:

@WardBrian
WardBrian enabled auto-merge November 5, 2026 21:33
@stan-buildbot

Copy link
Copy Markdown
Contributor
Name Old Result New Result Ratio Performance change( 1 - new / old )
stat_comp_benchmarks/benchmarks/low_dim_corr_gauss/low_dim_corr_gauss.stan 0.03 0.01 4.55 78.01% faster
stat_comp_benchmarks/benchmarks/irt_2pl/irt_2pl.stan 4.57 4.44 1.03 2.83% faster
stat_comp_benchmarks/benchmarks/pkpd/one_comp_mm_elim_abs.stan 19.11 17.93 1.07 6.17% faster
stat_comp_benchmarks/benchmarks/pkpd/sim_one_comp_mm_elim_abs.stan 0.24 0.23 1.05 4.99% faster
stat_comp_benchmarks/benchmarks/low_dim_gauss_mix_collapse/low_dim_gauss_mix_collapse.stan 8.86 8.37 1.06 5.54% faster
stat_comp_benchmarks/benchmarks/gp_regr/gp_regr.stan 0.11 0.1 1.09 8.06% faster
stat_comp_benchmarks/benchmarks/gp_regr/gen_gp_data.stan 0.02 0.02 1.08 7.59% faster
stat_comp_benchmarks/benchmarks/eight_schools/eight_schools.stan 0.06 0.05 1.08 7.12% faster
stat_comp_benchmarks/benchmarks/low_dim_gauss_mix/low_dim_gauss_mix.stan 2.85 2.73 1.04 4.23% faster
stat_comp_benchmarks/benchmarks/arK/arK.stan 1.65 1.57 1.05 4.62% faster
stat_comp_benchmarks/benchmarks/sir/sir.stan 88.75 85.64 1.04 3.51% faster
stat_comp_benchmarks/benchmarks/garch/garch.stan 0.47 0.44 1.08 7.69% faster
stat_comp_benchmarks/benchmarks/gp_pois_regr/gp_pois_regr.stan 2.38 2.23 1.07 6.19% faster
stat_comp_benchmarks/benchmarks/arma/arma.stan 0.3 0.28 1.05 5.11% faster
performance.compilation 247.56 240.61 1.03 2.81% faster
Mean result: 1.290738129545988

Jenkins Console Log
Jenkins Build Stages
Commit hash: 47d777ec364f286708b30c37d783a2c35acaa780

Machine information
Distributor ID:	Ubuntu
Description:	Ubuntu 20.04.3 LTS
Release:	20.04
Codename:	focal

CPU:

Architecture:                            x86_64
CPU op-mode(s):                          32-bit, 64-bit
Byte Order:                              Little Endian
Address sizes:                           46 bits physical, 48 bits virtual
CPU(s):                                  80
On-line CPU(s) list:                     0-79
Thread(s) per core:                      2
Core(s) per socket:                      20
Socket(s):                               2
NUMA node(s):                            2
Vendor ID:                               GenuineIntel
CPU family:                              6
Model:                                   85
Model name:                              Intel(R) Xeon(R) Gold 6248 CPU @ 2.50GHz
Stepping:                                7
CPU MHz:                                 1000.000
CPU max MHz:                             3900.0000
CPU min MHz:                             1000.0000
BogoMIPS:                                5000.00
L1d cache:                               1.3 MiB
L1i cache:                               1.3 MiB
L2 cache:                                40 MiB
L3 cache:                                55 MiB
NUMA node0 CPU(s):                       0,2,4,6,8,10,12,14,16,18,20,22,24,26,28,30,32,34,36,38,40,42,44,46,48,50,52,54,56,58,60,62,64,66,68,70,72,74,76,78
NUMA node1 CPU(s):                       1,3,5,7,9,11,13,15,17,19,21,23,25,27,29,31,33,35,37,39,41,43,45,47,49,51,53,55,57,59,61,63,65,67,69,71,73,75,77,79
Vulnerability Gather data sampling:      Mitigation; Microcode
Vulnerability Indirect target selection: Mitigation; Aligned branch/return thunks
Vulnerability Itlb multihit:             KVM: Mitigation: VMX unsupported
Vulnerability L1tf:                      Not affected
Vulnerability Mds:                       Not affected
Vulnerability Meltdown:                  Not affected
Vulnerability Mmio stale data:           Mitigation; Clear CPU buffers; SMT vulnerable
Vulnerability Old microcode:             Not affected
Vulnerability Reg file data sampling:    Not affected
Vulnerability Retbleed:                  Mitigation; Enhanced IBRS
Vulnerability Spec rstack overflow:      Not affected
Vulnerability Spec store bypass:         Mitigation; Speculative Store Bypass disabled via prctl
Vulnerability Spectre v1:                Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2:                Mitigation; Enhanced / Automatic IBRS; IBPB conditional; PBRSB-eIBRS SW sequence; BHI SW loop, KVM SW loop
Vulnerability Srbds:                     Not affected
Vulnerability Tsa:                       Not affected
Vulnerability Tsx async abort:           Mitigation; TSX disabled
Vulnerability Vmscape:                   Mitigation; IBPB before exit to userspace
Flags:                                   fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx pdpe1gb rdtscp lm constant_tsc art arch_perfmon pebs bts rep_good nopl xtopology nonstop_tsc cpuid aperfmperf pni pclmulqdq dtes64 monitor ds_cpl smx est tm2 ssse3 sdbg fma cx16 xtpr pdcm pcid dca sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand lahf_lm abm 3dnowprefetch cpuid_fault epb cat_l3 cdp_l3 intel_ppin ssbd mba ibrs ibpb stibp ibrs_enhanced fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid cqm mpx rdt_a avx512f avx512dq rdseed adx smap clflushopt clwb intel_pt avx512cd avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local dtherm ida arat pln pts hwp hwp_act_window hwp_epp hwp_pkg_req pku ospke avx512_vnni md_clear flush_l1d arch_capabilities

G++:

g++ (Ubuntu 9.4.0-1ubuntu1~20.04) 9.4.0
Copyright (C) 2019 Free Software Foundation, Inc.
This is free software; see the source for copying conditions.  There is NO
warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.

Clang:

clang version 10.0.0-4ubuntu1 
Target: x86_64-pc-linux-gnu
Thread model: posix
InstalledDir: /usr/bin

@WardBrian
WardBrian merged commit fa522ff into develop Nov 6, 2026
3 checks passed
@WardBrian
WardBrian deleted the fix-psis-zero-excesses branch November 6, 2026 00:52
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants