pith. sign in

arxiv: 2605.23134 · v1 · pith:VRAPZP3Tnew · submitted 2026-05-22 · 💻 cs.LG

Archimedean Copula Inference via Taylor-Mode AD

Cambridge Yang , Dongdong Li This is my paper

Pith reviewed 2026-05-25 05:00 UTC · model grok-4.3

classification 💻 cs.LG
keywords Archimedean copulasnested copulasautomatic differentiationcensoringsurvival analysismaximum likelihood estimationJAX
0
0 comments X

The pith

Polynomial powering of Taylor-mode AD computes exact nested Archimedean copula likelihoods and gradients under arbitrary censoring.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper presents a JAX framework that evaluates exact likelihoods and parameter gradients for nested Archimedean copulas with any nesting tree and any pattern of right-censoring. It replaces per-family hand-derived formulas with a single computation driven by polynomial powering of Taylor-mode automatic differentiation output. This works for both classical generators and user-defined neural ones. A sympathetic reader would care because it removes the barrier of custom derivations that has limited prior tools to low dimensions or simple nestings, enabling direct use on high-dimensional censored survival data.

Core claim

acopula evaluates exact nested-copula likelihoods and parameter gradients under arbitrary censoring masks in polynomial time. The mechanism is polynomial powering of Taylor-mode automatic differentiation output, which replaces per-family hand-derived partial Bell polynomial tables with a single differentiable computation that any user-defined generator can drive.

What carries the argument

polynomial powering of Taylor-mode automatic differentiation output, which replaces per-family hand-derived partial Bell polynomial tables with a single differentiable computation

If this is right

  • Enables per-variable censoring analysis on d=53 datasets such as MIMIC-IV ICU admissions using both classical and neural generators.
  • Supports hierarchical models on d=98 data such as an 11-sector S&P 500 returns model.
  • Allows family-agnostic censored maximum likelihood estimation across ten families, including five without prior implementations.
  • Delivers approximately 650 times speedup per density evaluation at d=35, with quadratic scaling to d=8000.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same powering approach could be tested on dependence structures outside the Archimedean class where similar polynomial identities arise.
  • Integration with other automatic-differentiation libraries beyond JAX would test whether the exactness property transfers without modification.
  • Application to even larger censoring patterns in longitudinal studies could reveal whether the polynomial-time bound remains practical in practice.

Load-bearing premise

The polynomial powering of Taylor-mode automatic differentiation output produces exact results for arbitrary nesting trees and user-defined generators without numerical instability or loss of differentiability.

What would settle it

Direct numerical comparison of likelihood values produced by the framework against analytically derived expressions for small nested copulas with known censoring masks.

Figures

Figures reproduced from arXiv: 2605.23134 by Cambridge Yang, Dongdong Li.

Figure 1
Figure 1. Figure 1: Nested Archimedean cop￾ula on six leaves. Nodes carry gener￾ators forming copula CDFs of their children; internal subtrees capture within-group dependence. Leaves are variables with marginal densities; the hatched leaf u3 is censored. A copula C : [0, 1]d → [0, 1] captures the dependence among d uniform random variables, independent of their marginal dis￾tributions [Sklar, 1959, Embrechts et al., 2002, Nel… view at source ↗
Figure 2
Figure 2. Figure 2: Wall-clock scaling. R only runs at 2-level K=5; aborts past d=40 on Clayton/Gumbel. By d=35 R is ∼650× slower. ACOPULA extends across topologies R cannot represent and reaches O(d 2 ) at d ≥ 6,400. A wall-clock comparison against R’s nacLL on matched single-thread CPU conditions at n=1 per evaluation shows ACOPULA growing polynomially in d while nacLL grows roughly exponentially in the number of sectors: b… view at source ↗
Figure 3
Figure 3. Figure 3: Head-to-head against R nacLL under the balanced M=K=⌈ √ d ⌉ topology, n=1 single￾evaluation timings, single-thread CPU, float64. R aborts at d≥40 for Clayton/Gumbel and on Frank for every d. ACOPULA extends to d=6,400 with the same O(d 2 ) asymptote as the fixed-K=5 topology ( [PITH_FULL_IMAGE:figures/full_fig_p035_3.png] view at source ↗
read the original abstract

No existing nested Archimedean copula tool handles all three of (a) arbitrary per-variable (right-)censoring in survival analysis, (b) arbitrary nesting trees, and (c) exact parameter gradients. Existing implementations handle only bivariate problems, low dimensional (i.e., $d \leq 10$) cases, two layers of nesting, or only hand-derived copula nestings. We present \textsc{acopula}, a JAX-native framework that, given any Archimedean generator -- classical or neural -- evaluates exact nested-copula likelihoods and parameter gradients under arbitrary censoring masks in polynomial time. The mechanism is polynomial powering of Taylor-mode automatic differentiation output, which replaces per-family hand-derived partial Bell polynomial tables with a single differentiable computation that any user-defined generator can drive. We conduct extensive simulations to verify the correctness of \textsc{acopula}. We then demonstrate (a) per-variable censoring on $85{,}229$ MIMIC-IV ICU admissions in high dimensions with $d{=}53$, fit by both classical Archimedean families and nested neural Archimedean copulas; (b) an 11-sector hierarchical model on S\&P~500 daily returns at $d{=}98$; (c) family-agnostic censored MLE across ten families, five of them with no prior implementation, on a retinopathy study; and (d) a ${\sim}650\times$ per-density speedup over R's \texttt{nacLL} at $d{=}35$, scaling quadratically to $d{=}8{,}000$.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 1 minor

Summary. The manuscript introduces acopula, a JAX-native framework for nested Archimedean copulas that, given any generator (classical or neural), computes exact likelihoods and parameter gradients under arbitrary nesting trees and per-variable censoring masks. The core mechanism is polynomial powering of Taylor-mode automatic differentiation output, which is claimed to replace hand-derived partial Bell polynomial tables with a single differentiable computation running in polynomial time. Correctness is asserted via extensive simulations, followed by demonstrations on high-dimensional censored data (MIMIC-IV d=53), financial returns (S&P 500 d=98), a retinopathy study across ten families, and a reported ~650x speedup over R's nacLL at d=35.

Significance. If the central mechanism of exact, stable, differentiable polynomial powering holds for arbitrary trees and user-defined generators, the work would enable flexible high-dimensional copula inference with neural generators and censoring, which existing tools do not support at scale. The reported applications and speedups indicate potential practical impact in survival analysis and finance if the exactness claim is substantiated.

major comments (1)
  1. [Abstract] Abstract: the central claim that polynomial powering of Taylor-mode AD output produces exact nested-copula likelihoods and gradients for arbitrary nesting trees and generators (replacing hand-derived partial Bell polynomials) is load-bearing, yet the provided text supplies no derivation, coefficient-array algebra, or analysis of floating-point accumulation or differentiability preservation under deep compositions.
minor comments (1)
  1. [Abstract] Abstract: the notation 85{,}229 appears to be a LaTeX artifact for 85,229 and should be rendered consistently.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for highlighting the load-bearing nature of the central claim. We agree that the abstract is too terse and that the main text (as submitted) does not contain a self-contained derivation of the polynomial-powering step, the coefficient-array algebra, floating-point analysis, or differentiability preservation. We will revise accordingly.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claim that polynomial powering of Taylor-mode AD output produces exact nested-copula likelihoods and gradients for arbitrary nesting trees and generators (replacing hand-derived partial Bell polynomials) is load-bearing, yet the provided text supplies no derivation, coefficient-array algebra, or analysis of floating-point accumulation or differentiability preservation under deep compositions.

    Authors: We accept the criticism. The submitted manuscript contains only a high-level description of the mechanism. In the revision we will (i) add a concise derivation of the polynomial-powering identity for arbitrary nesting trees in a new subsection of Section 3, including the explicit coefficient-array recurrence that replaces the partial Bell polynomials; (ii) include a short floating-point error analysis (forward-mode Taylor coefficients remain exact up to machine epsilon for the generator evaluations, with quadratic accumulation bounded by tree depth); and (iii) prove that the resulting likelihood and gradient remain differentiable with respect to both parameters and generator weights because the powering operation is a composition of analytic functions. These additions will be placed in the main text rather than the supplement. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation is self-contained

full rationale

The paper introduces a JAX-based computational method that applies Taylor-mode automatic differentiation followed by polynomial powering to evaluate nested Archimedean copula likelihoods and gradients for arbitrary generators and nesting trees. This replaces hand-derived partial Bell polynomials with a general AD-driven procedure. No equations or claims reduce by construction to fitted inputs, self-definitions, or load-bearing self-citations. The central mechanism is presented as a general technique whose correctness is checked via external simulations and real-data demonstrations rather than being presupposed by the method itself. The derivation chain therefore contains no circular steps.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review; no explicit free parameters, axioms, or invented entities are stated. The central claim rests on the unelaborated correctness of Taylor-mode AD powering for copula likelihoods.

pith-pipeline@v0.9.0 · 5823 in / 1127 out tokens · 29829 ms · 2026-05-25T05:00:28.232386+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

44 extracted references · 44 canonical work pages · 2 internal anchors

  1. [1]

    Pair-copula constructions of multiple dependence

    Kjersti Aas, Claudia Czado, Arnoldo Frigessi, and Henrik Bakken. Pair-copula constructions of multiple dependence. Insurance: Mathematics and Economics, 44 0 (2), 2009

    work page 2009

  2. [2]

    A Geometric Theory of Higher-Order Automatic Differentiation

    Michael Betancourt. A geometric theory of higher-order automatic differentiation. arXiv preprint arXiv:1812.11592, 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  3. [3]

    Taylor-mode automatic differentiation for higher-order derivatives in JAX

    Jesse Bettencourt, Matthew James Johnson, and David Duvenaud. Taylor-mode automatic differentiation for higher-order derivatives in JAX . In Program Transformations for ML Workshop at NeurIPS, 2019

    work page 2019

  4. [4]

    Generalized autoregressive conditional heteroskedasticity

    Tim Bollerslev. Generalized autoregressive conditional heteroskedasticity. Journal of Econometrics, 31 0 (3), 1986

    work page 1986

  5. [5]

    JAX : Composable transformations of P ython+ N um P y programs, 2018

    James Bradbury, Roy Frostig, Peter Hawkins, Matthew James Johnson, Chris Leary, Dougal Maclaurin, George Necula, Adam Paszke, Jake Vander P las, Skye Wanderman-Milne, and Qiao Zhang. JAX : Composable transformations of P ython+ N um P y programs, 2018. URL http://github.com/jax-ml/jax

    work page 2018

  6. [6]

    Hierarchical K endall copulas: Properties and inference

    Eike Christian Brechmann. Hierarchical K endall copulas: Properties and inference. Canadian Journal of Statistics, 42 0 (1), 2014

    work page 2014

  7. [7]

    Xinyuan Chen, Yiwei Li, and Qian M. Zhou. Analysis of multivariate event times under informative censoring using vine copula. arXiv preprint arXiv:2502.20608, 2025

    work page arXiv 2025

  8. [8]

    Advanced Combinatorics: The Art of Finite and Infinite Expansions

    Louis Comtet. Advanced Combinatorics: The Art of Finite and Infinite Expansions. D. Reidel, 1974

    work page 1974

  9. [9]

    Correlation and dependence in risk management: properties and pitfalls

    Paul Embrechts, Alexander McNeil, and Daniel Straumann. Correlation and dependence in risk management: properties and pitfalls. In M. A. H. Dempster, editor, Risk Management: Value at Risk and Beyond, pages 176--223. Cambridge University Press, 2002

    work page 2002

  10. [10]

    Analysis of Survival Data with Dependent Censoring: Copula-Based Approaches

    Takeshi Emura and Yi-Hau Chen. Analysis of Survival Data with Dependent Censoring: Copula-Based Approaches. Springer, 2018

    work page 2018

  11. [11]

    Krishnan

    Ali Hossein Gharari Foomani, Michael Cooper, Russell Greiner, and Rahul G. Krishnan. Copula-based deep survival models for dependent censoring. In Conference on Uncertainty in Artificial Intelligence, 2023

    work page 2023

  12. [12]

    A semiparametric estimation procedure of dependence parameters in multivariate families of distributions

    Christian Genest, Kilani Ghoudi, and Louis-Paul Rivest. A semiparametric estimation procedure of dependence parameters in multivariate families of distributions. Biometrika, 82 0 (3), 1995

    work page 1995

  13. [13]

    Evaluating Derivatives: Principles and Techniques of Algorithmic Differentiation

    Andreas Griewank and Andrea Walther. Evaluating Derivatives: Principles and Techniques of Algorithmic Differentiation. SIAM, 2008

    work page 2008

  14. [14]

    Nicholas J. Higham. Accuracy and Stability of Numerical Algorithms. SIAM, 2002

    work page 2002

  15. [15]

    Sampling A rchimedean copulas

    Marius Hofert. Sampling A rchimedean copulas. Computational Statistics & Data Analysis, 52 0 (12), 2008

    work page 2008

  16. [16]

    Sampling Nested A rchimedean Copulas: With Applications to CDO Pricing

    Marius Hofert. Sampling Nested A rchimedean Copulas: With Applications to CDO Pricing . PhD thesis, Ulm University, 2010

    work page 2010

  17. [17]

    Densities of nested A rchimedean copulas

    Marius Hofert and David Pham. Densities of nested A rchimedean copulas. Journal of Multivariate Analysis, 118, 2013

    work page 2013

  18. [18]

    Marius Hofert, Martin M \"a chler, and Alexander J. McNeil. Likelihood inference for A rchimedean copulas in high dimensions under known margins. Journal of Multivariate Analysis, 110, 2012

    work page 2012

  19. [19]

    Marius Hofert, Martin M \"a chler, and Alexander J. McNeil. A rchimedean copulas in high dimensions: Estimators and numerical challenges motivated by financial applications. Journal de la Soci \'e t \'e Fran c aise de Statistique , 154 0 (1), 2013

    work page 2013

  20. [20]

    Huster, Ron Brookmeyer, and Steven G

    William J. Huster, Ron Brookmeyer, and Steven G. Self. Modelling paired survival data with covariates. Biometrics, 45 0 (1), 1989

    work page 1989

  21. [21]

    Harry Joe and James J. Xu. The estimation method of inference functions for margins for multivariate models. Technical Report 166, Department of Statistics, University of British Columbia, 1996

    work page 1996

  22. [22]

    Alistair E. W. Johnson, Lucas Bulgarelli, Lu Shen, Alvin Gayles, Ayad Shammout, Steven Horng, Tom J. Pollard, Sicheng Hao, Benjamin Moody, Brian Gow, Li wei H. Lehman, Leo A. Celi, and Roger G. Mark. MIMIC-IV , a freely accessible electronic health record dataset. Scientific Data, 10 0 (1), 2023

    work page 2023

  23. [23]

    Copulas.jl: A fully distributions.jl-compliant copula package

    Oskar Laverny and Santiago Jimenez. Copulas.jl: A fully distributions.jl-compliant copula package. Journal of Open Source Software, 9 0 (94): 0 6189, 2024

    work page 2024

  24. [24]

    Joan Hu, Mary L

    Dongdong Li, X. Joan Hu, Mary L. McBride, and John J. Spinelli. Multiple event times in the presence of informative censoring: modeling and analysis by copulas. Lifetime Data Analysis, 26 0 (3), 2020

    work page 2020

  25. [25]

    Joan Hu, and Rui Wang

    Dongdong Li, X. Joan Hu, and Rui Wang. Evaluating association between two event times with observations subject to informative censoring. Journal of the American Statistical Association, 118 0 (542), 2023

    work page 2023

  26. [26]

    Automatic functional differentiation in JAX

    Min Lin. Automatic functional differentiation in JAX . In International Conference on Learning Representations, 2024

    work page 2024

  27. [27]

    Zico Kolter

    Chun Kai Ling, Fei Fang, and J. Zico Kolter. Deep A rchimedean copulas. In Advances in Neural Information Processing Systems, 2020

    work page 2020

  28. [28]

    Discrete copula diffusion

    Anji Liu, Oliver Broadrick, Mathias Niepert, and Guy Van den Broeck. Discrete copula diffusion. In International Conference on Learning Representations, 2025 a

    work page 2025

  29. [29]

    HACSurv : A hierarchical copula-based approach for survival analysis with dependent competing risks

    Xin Liu, Weijia Zhang, and Min-Ling Zhang. HACSurv : A hierarchical copula-based approach for survival analysis with dependent competing risks. In International Conference on Artificial Intelligence and Statistics, 2025 b

    work page 2025

  30. [30]

    Alexander J. McNeil. Sampling nested A rchimedean copulas. Journal of Statistical Computation and Simulation, 78 0 (6), 2008

    work page 2008

  31. [31]

    McNeil and Johanna Ne s lehov \'a

    Alexander J. McNeil and Johanna Ne s lehov \'a . Multivariate A rchimedean copulas, d -monotone functions and _1 -norm symmetric distributions. The Annals of Statistics, 37 0 (5B), 2009

    work page 2009

  32. [32]

    rvinecopulib: High Performance Algorithms for Vine Copula Modeling, 2023

    Thomas Nagler and Thibault Vatter. rvinecopulib: High Performance Algorithms for Vine Copula Modeling, 2023. URL https://CRAN.R-project.org/package=rvinecopulib. R package, CRAN

    work page 2023

  33. [33]

    Roger B. Nelsen. An Introduction to Copulas. Springer, 2006

    work page 2006

  34. [34]

    Generative A rchimedean copulas

    Yuting Ng, Ali Hasan, Khalil Elkhalil, and Vahid Tarokh. Generative A rchimedean copulas. In Uncertainty in Artificial Intelligence, 2021

    work page 2021

  35. [35]

    Hierarchical A rchimedean copulae: The HAC package

    Ostap Okhrin and Alexander Ristig. Hierarchical A rchimedean copulae: The HAC package. Journal of Statistical Software, 58 0 (4), 2014

    work page 2014

  36. [36]

    Composable Effects for Flexible and Accelerated Probabilistic Programming in NumPyro

    Du Phan, Neeraj Pradhan, and Martin Jankowiak. Composable effects for flexible and accelerated probabilistic programming in N um P yro. arXiv preprint arXiv:1912.11554, 2019

    work page internal anchor Pith review Pith/arXiv arXiv 1912

  37. [37]

    You only linearize once: Tangents transpose to gradients

    Alexey Radul, Adam Paszke, Roy Frostig, Matthew Johnson, and Dougal Maclaurin. You only linearize once: Tangents transpose to gradients. Proceedings of the ACM on Programming Languages, 7 0 (POPL), 2023

    work page 2023

  38. [38]

    On the construction of nested A rchimedean copulas for d -monotone generators

    Mohsen Rezapour. On the construction of nested A rchimedean copulas for d -monotone generators. Statistics & Probability Letters, 101, 2015

    work page 2015

  39. [39]

    Fonctions de r \'e partition \`a n dimensions et leurs marges

    Abe Sklar. Fonctions de r \'e partition \`a n dimensions et leurs marges. Publications de l'Institut de Statistique de l'Universit \'e de Paris , 8, 1959

    work page 1959

  40. [40]

    Simplified pair copula constructions---limitations and extensions

    Jakob St \"o ber, Harry Joe, and Claudia Czado. Simplified pair copula constructions---limitations and extensions. Journal of Multivariate Analysis, 119, 2013

    work page 2013

  41. [41]

    CopulaCenR : Copula based regression models for bivariate censored data in R

    Tao Sun and Ying Ding. CopulaCenR : Copula based regression models for bivariate censored data in R . The R Journal, 12 0 (1), 2020

    work page 2020

  42. [42]

    Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan T

    Matthew Tancik, Pratul P. Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan T. Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. In Advances in Neural Information Processing Systems, 2020

    work page 2020

  43. [43]

    A statistical distribution function of wide applicability

    Waloddi Weibull. A statistical distribution function of wide applicability. Journal of Applied Mechanics, 18 0 (3), 1951

    work page 1951

  44. [44]

    Ming Zheng and John P. Klein. Estimates of marginal survival for dependent competing risks based on an assumed copula. Biometrika, 82 0 (1), 1995

    work page 1995