Changes in version 0.4.0 (2026-11-07)                  

  - CoxDeviance$information_derivative(linear_predictor, direction,
    sample_weight): the derivative of the information in a direction of
    the linear predictor, d/dt I(eta + t direction), as a symmetric
    matrix-free operator (CoxInformationDerivative, with %*%, t(), dim()
    and as.matrix()), O(n) per product: for the derivatives of log
    determinants, as in REML.
  - The C++ library is header-only: inst/include has coxdev.h and its
    implementation coxdev_impl.h, all inline, so another package can
    compile against it with LinkingTo: coxdev and #include <coxdev.h>.
  - glmnet is no longer in Suggests: the comparisons with it (and the
    vignettes' use of it) are removed, as glmnet will depend on coxdev
    (LinkingTo: coxdev). The comparisons with survival remain.
  - More tests, mirroring the Python package's: the residuals, the
    baseline hazard, the expected events and the risk-set variances
    against direct computations, with zero weights; the C-index against
    the count of its definition and glmnet::Cindex; strata (a model per
    stratum, the block diagonal information, labels, strata of unequal
    sizes against coxph); zero weights as dropped observations, for
    every kind of ties; the caching of results; the problematic case of
    tests/test_bad.py; the moments of the "fisher" draws; edge cases (no
    failures, one failure, everyone failing at once, entries at failure
    times, strata without failures, extreme linear predictors) and
    finite differences of the derivatives under every option; and the
    values of the Python package, draws of sample_score() included, for
    a fixed data set.

                        Changes in version 0.3.0                        

The first release of the R package. It runs the C++ library of the
Python package coxdev 0.3.0 and has the same interface:
CoxDeviance$new(...) in place of CoxDeviance(...).

  - CoxDeviance$new() takes the response as vectors or as a Surv object,
    with start times, strata (any labels), sample weights, and Breslow's
    or Efron's ties.
  - compute(): the deviance, its gradient and the diagonal of its
    Hessian in the linear predictor; information(): the information as a
    matrix-free operator.
  - sample_score() and wild_score(): draws of the score (multinomial,
    Bernoulli, Fisher's noncentral hypergeometric, wild).
  - score_residuals(), schoenfeld_residuals() and risk_set_variance(), a
    column at a time for a dense or sparse design.
  - baseline_hazard(), expected_events() and survival_at_exit().
  - c_index(): Harrell's concordance index, in O(n log n).

All agree with the survival package, and the vignettes match the Python
notebooks.