Skip to content

log_modified_bessel_first_kind has zero derivatives in the order at v = 0 and v = 1 #3510

Open

Description

@jachymb

AI use disclosure: This was written with the help of Claude Code (Opus 5.5).

Description

For v == 0 and v == 1, log_modified_bessel_first_kind(v, z) evaluates Boost's fits for $I_0$ and $I_1$, which do not depend on v. When v is an autodiff variable, every derivative with respect to v is therefore 0, in reverse and forward mode. The exact first derivative is $-K_0(z)/I_0(z)$ at v = 0 (decaying like $e^{-2z}$) and about $-1/z$ at v = 1 for large z. The source comments acknowledge this ("WARNING: will not autodiff for v = 0 correctly").

A sampler, optimizer or diagnose can start exactly there: with init=0, a real<lower=0> order starts at exp(0) = 1 and a real<lower=-1> order at 0.

do_lkj_constant, the normalizing constant of lkj_corr_lpdf, lkj_corr_cholesky_lpdf and lkj_cov_lpdf, has the same defect at eta == 1.0: develop drops the η-derivatives of the constant there, so the η-gradient keeps only the data term and the second η-derivative is 0. Open PR #3434 fixes this as part of its rewrite: on its head (35435f5), the first and second η-derivatives at η = 1 are within 3e-16 of the exact values (K = 2). Its tests have no η = 1 point.

Possible fix: take the two shortcuts only when v is not autodiff (!is_autodiff_v<T1> &&). The test below then passes, and an autodiff order goes through the general power series.

Example

New file test/unit/math/rev/fun/log_modified_bessel_first_kind_test.cpp:

#include <stan/math/rev.hpp>
#include <test/unit/math/rev/util.hpp>
#include <gtest/gtest.h>

TEST_F(AgradRev, log_modified_bessel_first_kind_order_gradient) {
  using stan::math::var;
  auto d_dv = [](double v_val) {
    var v = v_val;
    stan::math::log_modified_bessel_first_kind(v, 0.5).grad();
    double g = v.adj();
    stan::math::recover_memory();
    return g;
  };
  // d/dv log I_v(0.5) from mpmath (50 digits); at v = 0 it is -K_0 / I_0
  EXPECT_NEAR(-0.86923697789771742858, d_dv(0), 1e-12);
  EXPECT_NEAR(-1.8244901560201558681, d_dv(1), 1e-12);
}

It fails on develop:

[ RUN      ] AgradRev.log_modified_bessel_first_kind_order_gradient
d_dv(0) evaluates to 0
d_dv(1) evaluates to 0
[  FAILED  ] AgradRev.log_modified_bessel_first_kind_order_gradient

Expected Output

The order gradient at z = 0.5 is −0.86923697789771743 at v = 0 and −1.8244901560201559 at v = 1 (mpmath).

Current Version:

v5.4.0 (develop at 1f83dc5)

Activity

  1. jachymb commented on Nov 13, 2026

    @jachymb
    ContributorAuthor

    Somewhat related to, but technically independent of #3512

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions