                 Changes in version 0.2.3 (2026-10-07)                  

  - Use pivoted QR preconditioning for the gauge linear program to
    retain accuracy for nearly collinear, full-rank influence paths.
    Check the returned direction for feasibility and clarify solver
    failure messages.
  - Standardize the GMM first-order-condition warning by moment
    variation; record the residual and numerical tolerance without
    treating the check as a test of asymptotic assumptions.
  - Clarify the registry's MIT license, installation instructions,
    scalar reference memory costs, and the target dimension used by HAC
    adjustment.
  - Repair working-paper links, author initials and PDF guide output
    wrapping.
  - Add regression tests for ill-conditioning, pivot/direction mapping,
    test-region consistency and optimizer-based GMM estimates.

                        Changes in version 0.2.2                        

  - Update the main-method citation to Cambridge Working Papers in
    Economics No. 2678 and Janeway Institute Working Paper No. 2637,
    with the public URL.
  - Prepare the first CRAN submission; no changes to numerical methods
    or APIs.

                        Changes in version 0.2.1                        

  - Bind reusable normalizers to their influence contributions and
    profile; reject mismatched data and tuning supplied with a prebuilt
    object.
  - Apply calendar-time restrictions to direct HAC, fixed-b and EWC
    constructors as well as the common interface.
  - Check quadratic joint regions and intervals for singularity,
    assessing dependence after coordinate scaling so changes of units
    remain usable.
  - Correct the returned LDL transformation for a requested coordinate
    order, reject fractional permutations, and improve LDL scaling
    checks.
  - Validate fixed-b reference bandwidth/grid consistency and reference
    arguments.
  - Report separately selected tuning for each marginal contrast.
  - Stabilize quadratic spectral weights near zero; clarify bandwidth
    versus lag, the asymptotic EWC reference, plot labels and
    installation instructions.

                        Changes in version 0.2.0                        

Local review candidate. Not released, not submitted.

Five further inference methods

Inference is no longer restricted to the affine-equivariant
adjusted-range increment hull, which remains the default and is
unchanged. Five methods can now be applied to the same fit, the same
influence contributions, the same null value and the same level, through
a method argument on aersn_test(), confint(), aersn_contrast(),
aersn_region() and plot(), and side by side through the new
aersn_compare():

  - "ldl", componentwise adjusted ranges after lag-zero partial
    prewhitening by an LDL factorization (aersn_ldl_normalizer(),
    aersn_ldl());
  - "shao", quadratic self-normalization, with calendar-time or
    variance-profile integration (aersn_shao_normalizer());
  - "hac", kernel long-run covariance estimation with the Bartlett,
    Parzen and quadratic spectral kernels, four bandwidth rules,
    explicit bandwidths or lag truncations, optional VAR(1) prewhitening
    and an optional finite-sample multiplier (aersn_hac_lrv());
  - "fixedb", Bartlett fixed-b (aersn_fixed_b_normalizer());
  - "ewc", the equal-weighted cosine method (aersn_ewc_lrv()).

aersn_normalizer() builds the normalizer of any method and reports the
tuning values actually used, including bandwidths selected from the data
and fixed-b fractions rounded to an integer lag truncation.

Reference laws

aersn_reference() gains statistic and args arguments and simulates the
matched-grid law of the componentwise, quadratic and fixed-b statistics
as well as the increment-hull gauge. aersn_parametric_reference()
provides the chi-squared law used with HAC and the scaled F law used
with EWC. Cache keys, and the check that a statistic is combined with
its own law, now include the method and every tuning value that changes
the distribution, so a law simulated for one method or one bandwidth
fraction cannot be used with another. The scalar componentwise law
reuses the scalar hull draws, so the two agree exactly.

Verified identities

The test suite checks, among others, that the Bartlett fixed-b estimate
at b = 1 is exactly twice the quadratic self-normalizer and that the two
tests then agree on common reference draws; that the HAC estimate and
its Wald statistic reproduce sandwich::vcovHAC() to machine precision
for all three kernels, with and without prewhitening; that the Andrews
bandwidths for the Parzen and quadratic spectral kernels stand in the
ratio 2.6614/1.3221; that fixed-b at a realized fraction m/n equals the
Bartlett HAC estimate at lag truncation m - 1, the two conventions
differing by one; and that the componentwise statistic is unchanged by
the choice of divisor in the lag-zero covariance and by the use of the
Cholesky factor in place of the unit lower triangular one.

Other changes

  - nu has its own argument in the inference functions. It is a prefix
    of null, so R's partial matching would otherwise have sent the
    number of cosine terms to the null value.
  - summary() reports why intervals are unavailable instead of omitting
    them silently.
  - Confidence regions keep the geometry of their method: a convex
    polygon for the increment hull, an ellipsoid for the other five.
    aersn_projection(), aersn_slice(), aersn_contains() and plot() work
    for both.
  - New vignette, Comparing inference methods on one estimate.
  - sandwich is a new dependency, used for the Andrews and Newey-West
    bandwidth rules.

Scope

The structural-break test of Hong, Linton, McCabe, Sun and Wang (2024)
and the autocorrelation tests of Sun, Zhu and Linton (2025) remain
outside the package. No method here is an adjusted-range version of
either: substituting a different normalizer into a published statistic
would need its own theory and reference law.

                        Changes in version 0.1.1                        

Review fixes. See dev/FIXES_0.1.1.md for the full list: reference-cache
keys, missing-value handling in aersn_lm(), degenerate profiles and
failed reference draws, quantile-table and parameter-selection
validation, one-sided tests, Monte Carlo uncertainty, numerical scaling,
and plotting. Authorship was completed to the five manuscript authors
with Jiajing Sun as maintainer.

                        Changes in version 0.1.0                        

Initial development version.

  - Core construction: centered influence path (aersn_path()),
    increment-hull self-normalizer (aersn_hull()), gauge by linear
    programming (aersn_gauge()), support function (aersn_support()).
  - Inference: aersn(), aersn_test(), confint() with simultaneous, joint
    and marginal intervals, aersn_contrast(), aersn_region(),
    aersn_contains(), two-dimensional projections and slices, plotting.
  - Reference distributions: closed-form scalar law, matched-grid Monte
    Carlo references with reproducible seeds, batch Monte Carlo standard
    errors, caching, and the verified registry aersn_registry.
  - Model interfaces: aersn_mean(), aersn_lm(), aersn_gmm(),
    aersn_mle(), and aersn_target().
