NEWS

---

HDElliptical 0.1.3 (2026-09-14)

• Made two numerical-boundary tests portable to Linux ARM64 by allowing either mathematically degenerate trace estimate to trigger the documented no-repair error and by using an exactly representable cross-zero fixture.
• Relaxed two chi-square p-value comparisons to a still stringent 1e-12 tolerance so the tests do not depend on extended-precision intermediates.

HDElliptical 0.1.2 (2026-09-09)

• Expanded Authors@R to identify all package contributors and copyright holders, including Dan Zhuang's scoped SEMC and GEMcluster contribution.
• Clarified package, SEMC, reference-implementation, and test-fixture provenance in inst/COPYRIGHTS and the affected source headers.
• Corrected two package-original Chapter 4 C++ files whose generic GPL SPDX headers conflicted with their intended MIT licensing.

HDElliptical 0.1.1

• Replaced README links to build-excluded audit and benchmark files with public GitHub URLs so they remain resolvable from the CRAN source package.
• Added an author-year citation and public source URL to the package description for CRAN submission metadata.

HDElliptical 0.1.0

Chapter 1 foundations

• Added Rcpp implementations of spatial signs, the modified-Weiszfeld spatial median, empirical spatial ranks, the multivariate spatial Kendall matrix, Tyler's fixed-point estimator, and the Hettmansperger-Randles iteration.
• Added R interfaces for SSCM and spatial-rank covariance matrices, ACG likelihood evaluation, shape normalization, and elliptical simulation.
• Added tests of defining equations and translation, scale, orthogonal, and affine equivariance properties.
• Hardened directional and shape calculations against overflow and underflow, including finite inputs near the limits of double precision.
• Changed iterative convergence certificates to use the defining estimating equation residuals, preventing false convergence near an observation.
• Defined the one-dimensional HR interface as the ordinary sample median with unit shape and an explicit not-applicable joint-shape diagnostic.

Chapter 2 location tests

• Added exact one- and two-sample Hotelling tests and fixed-dimensional spatial sign, signed-rank, and pooled spatial-rank tests.
• Added Rcpp implementations of the Srivastava--Du, Park--Ayyala, Bai--Saranadasa, Chen--Qin, and Srivastava--Katayama--Kano high-dimensional mean tests.
• Added the Cai--Liu--Xia precision-adjusted maximum test with deliberately separate oracle, supplied-feasible, and adaptive-threshold precision paths. The feasible path uses transformed within-group empirical variances rather than the diagonal of an estimated precision matrix.
• Added the Wang--Peng--Li one-sample high-dimensional spatial-sign test with its literal leave-two-out cross-validation variance estimator. Exact zero signs follow U(0) = 0; invalid variance estimates are not repaired.
• Added the Feng--Sun scalar-invariant one-sample spatial-sign test with literal pair-specific leave-two-out location/diagonal fits and the primary paper's feasible trace calibration omitted from the short book presentation.
• Added the Huang--Liu--Zhou--Feng two-sample inverse-norm sign test (tINST) with the paper's feasible cross statistic, observation-specific leave-one- out nuisance fits, published variance decomposition, and explicit iteration stability versus score-residual diagnostics.
• Added the Feng--Zou--Wang--Zhu scale-invariant high-dimensional Behrens--Fisher test with the primary paper's asymptotic centering and exact leave-four-out / two-plus-two leaveout trace estimators.
• Added the common-covariance Composite T2 test with paper-greedy correlation blocks, a full two-plus-two leaveout statistic, and the original first-group leave-four-out trace calibration.
• Added the Feng--Liu--Ma one-sample INST procedure with null-centered endpoints, delete-two nuisance fits, and the supplement's direct 2 * n^(-4) feasible variance.
• Added the Feng--Zou--Wang two-sample multivariate-sign test with crossed leave-one-out locations, full/leave-one-out diagonal HR fits, and the primary paper's Proposition 2 feasible trace and variance calibration.
• Added the Li--Wang--Zou simpler two-sample spatial-sign test with two full-sample diagonal HR fits, the exact feasible bias correction, and the published distinction between ordered-pair trace denominators and the outer squared-sample-size variance factors.
• Added the Feng--Zhang--Liu high-dimensional two-sample spatial-rank test with its feasible leave-four/leave-two trace calibration. The API exposes both a coherent geometric scale identification and the article's literal trace-normalized finite-sample definition.
• Added the Xu--Lin--Wei--Pan analytical aSPU test. Its default finite-power scores reproduce the primary paper's raw sample-mean differences, while an explicit book_studentized path implements the book's scale-invariant standardized-coordinate variant; both use deterministic odd/even Gaussian and maximum extreme-value calibration without paper-specific simulations.
• Added the scaled spatial median, the spatial-sign MAX test, and its MAXSUM Cauchy combination with the existing Feng--Sun SUM component. The API exposes fixed-point stability and estimating-equation residuals separately.
• Added the Yan--Zhao--Feng weighted scaled spatial median and general-m weighted MAX/MAXSUM tests. The m = -1 SUM path is locked component by component to the existing INST implementation, and the published half-power, endpoint-index, and radial-moment typos are corrected explicitly.
• Added Zhang--Feng one- and two-sample adaptive rank tests with max, squared- rank sum, and equal-weight Cauchy components. Sum calibration requires a supplied positive long-run variance or an explicitly selected Ouyang--Parzen lag; there is no hidden bandwidth or variance floor.
• Added generic_weighted_hr_location() and the deliberately oracle oracle_weighted_sign_sum_test() for the formula-complete generic weighted layer. Calibration is reported only from an explicit null_sd, or supplied nu2 and trace_R2; no feasible p-value is guessed.
• Added zhang_zhou_guo_one_sample_test(), zhang_zhou_guo_paired_test(), and zhang_zhou_guo_linear_hypothesis_test() for normal-reference one-sample means, rowwise paired differences, and same-unit linear hypotheses. All three wrappers reuse the compiled centred-U-statistic and unbiased-trace kernel, and deterministic tests lock the appendix v + 4 denominator, wrapper transformations, primal/dual agreement, invariances, and no-repair errors.
• Added zhang_zhu_zhang_two_sample_test(), the Zhang--Zhu--Zhang scale-invariant two-sample normal-reference test, with crossed covariance weights, bias-corrected trace calibration, and distinct published-threshold and unadjusted degrees-of-freedom paths. Deterministic formula, primal/dual, invariance, extreme-tail, and boundary checks cover the compiled implementation.
• Added wang_xu_approx_randomization_test() with the Wang--Xu full-sample CQ statistic and adjacent within-group half-difference reference sample. The compiled kernel supports exhaustive small-sample sign enumeration and a reproducible counter-based Monte Carlo mode; deterministic tests lock odd- row discard, tie handling, exact tails without +1, Monte Carlo tails with +1, seed isolation, and no-repair boundaries.
• Added feng_wang_pdq_two_sample_test(), the Feng--Wang PDQ two-sample spatial-sign test, with empirical pairwise-difference U-quantiles, standardized spatial medians, feasible bias/variance matrices, and Rademacher wild-bootstrap calibration. Literal small-sample and fixed-counter tests verify the compiled statistic, bootstrap, dual finite-B decision conventions, invariances, zero directions, extreme units, and degeneracy contracts.
• Added zhao_feng_strongcorr_sign_test(), the Zhao--Feng strong-correlation one-sample spatial-sign test, with Rademacher or Gaussian wild-bootstrap multipliers. The compiled pair-sum kernel and deterministic tests preserve the distinct null and fitted centerings, multiplier-dependent diagonal, common-tau cancellation, finite-B decisions, RNG contract, invariances, and zero/no-repair cases.
• Added elliptical_regularized_hotelling_test() and elliptical_regularized_hotelling_cauchy_test() for fixed-ridge and Cauchy-grid Feng--Zhou--Wang elliptical regularized Hotelling testing. One economy-SVD compiled kernel evaluates the feasible companion-matrix centre and variance for every ridge; deterministic tests cover literal margins, the p > n complement route, small positive ridges, stable signed-log Cauchy aggregation, invariances, extreme scales, and explicit failures. The paper's unspecified simulation-program Bartlett modification is not guessed.
• Marked ch2_location.tex:2438-2451 structured-correlation combinations as the sole blocked Chapter 2 item. A legally accessible primary full text has not been obtained and the book contains no executable formulas, so no API is fabricated pending source access.
• Retained randomization and wild bootstrap only where they are intrinsic test calibrations. These callable calibrations are not paper simulation replications; no paper-specific Monte Carlo grids, size/power scenarios, tables, reproduction scripts, or data were added.
• Stabilized CLX group centering and adaptive covariance moments under large common translations and representable extreme global scales.
• Kept the original common-covariance Bai--Saranadasa statistic separate from the unequal-covariance Chen--Qin U-statistic, correcting their conflation in the current book draft.
• Added stable primal/dual Gram calculations, leave-out kernels, scaled diagnostic components, and independent formula/invariance/degeneracy tests.

Chapter 3 matrix estimation and testing

• Added six classical covariance, sphericity, and spatial sign/rank shape tests, with the primary Mauchly, John, and Nagao finite-sample factors.
• Added Bickel--Levina, Rothman--Levina--Zhu, and Cai--Liu covariance thresholding plus supplied-factor POET, together with five high-dimensional Gaussian/light-tail covariance tests.
• Added four elliptical sign/rank/adaptive sphericity tests and two SSCM/spatial-rank proportionality tests with their feasible bias and variance calibration.
• Added certificate-checked spatial-sign SCLIME and SGLASSO precision estimation and exact support thresholding. Failed feasibility, KKT, or positive-definiteness checks are returned as failures rather than repaired.
• Added ec2_covariance() for the formula-complete convex-l1 EC2 branch, gaussian_graphical_lasso() with an off-diagonal l1 penalty, and clime_precision() with primary smaller-absolute-value symmetrization. Each reports strict feasibility/KKT or primal-dual certificates without hidden ridge, floor, pseudoinverse, or SPD repair. Adaptive/MC+ EC2, SCIO, and scaled-lasso remain review-only because their distinct contracts are not closed by the short review passages.
• Added tensor spatial-sign precision estimation and graph thresholding with column-major Kolda unfolding, the published pilot switch, and modewise objective/KKT diagnostics.
• Added the Ollila--Raninen shrinkage covariance estimator, regularized SSCM, BASIC, BASICS, and linear covariance pooling with self-contained quadrature, bounded inversion, and pooling optimisation.
• Added the primary joint high-dimensional HR fixed point, which bands the standardised-sign SSCM within every iteration and reports map, score, convergence, conditioning, and pilot diagnostics.
• Added elliptical factor-number selection, spatial-sign POET, one-step Tyler POET, and sparse factor precision reconstruction with explicit factor count and tuning constants.
• At the earlier Chapter 1--3 gate, generation produced 107 Rd files. That source package passed all examples, 35 test files, both vignettes, and R CMD check --no-manual with zero errors and warnings; the only NOTE is the existing Windows linked-symbol scan.
• Added no paper-specific simulation grids, size/power scenarios, result tables, reproduction scripts, or simulation data. Tests are deterministic formula, invariance, solver-certificate, and numerical-boundary checks.

Chapter 4 other high-dimensional tests

• Added 33 callable APIs spanning unconditional and conditional alpha tests, mutual-fund FDR, change-point testing and ERHT-WBS segmentation, white-noise tests, a radial-directional diagnostic, and Gaussian/panel/vector independence testing.
• Completed the formula-closed Chapter 4 layer with weighted_spatial_sign_alpha_oracle_test(), zhao_chen_zi_inst_alpha_test(), book_gaussian_alpha_cauchy_test(), conditional_factor_wald_test(), and wang_liu_feng_vector_u_independence_test(). The separate mutual-independence max--sum studentization remains source-blocked rather than being replaced by the adjacent exact rank-U vector method.
• Corrected the book-level DMS, spatial-sign change-point, alpha-test, conditional-residual, white-noise variance, panel-Gumbel, and serial-panel formula conflations against the primary sources; every software choice is exposed in returned diagnostics.
• Added strict nuisance-fit, covariance, determinant, variance, solver, and seed contracts. Invalid inputs or uncertified fits fail explicitly rather than receiving an implicit ridge, floor, pseudoinverse, or nearest-PD repair.
• Retained bootstrap, randomization, permutation, and Gaussian-process draws only where they define a method's own calibration. No paper simulation design, size/power grid, empirical table, reproduction script, or data set was added.
• At the earlier Chapter 1--4 gate, generation produced 113 exports and 147 Rd files. That source tarball passed 41 test files, 426 named test blocks, all examples, both vignettes, and R CMD check --no-manual with zero errors and warnings; the only NOTE is the existing Windows linked-symbol scan.

Chapter 5 classification

• Added a common hd_classifier_fit contract and 23 callable APIs covering oracle/classical LDA and QDA, independence and FAIR rules, thresholded and direct sparse LDA, sparse Gaussian/spatial-sign QDA, certified plug-in rules, and robust GQDA/HR-GQDA.
• Added strict vector- and matrix-Dantzig, lasso, and interaction-loss solver certificates. Tuning paths, priors, score orientation/scale, determinant signs, SPD status, feasibility, duality, and KKT residuals are retained in every fit; failed fits are not silently repaired or made predictable.
• Corrected the book's unequal-prior elliptical shortcut, DSDA coding/objective, Jiang versus Cai--Zhang attribution, explicit ridge-operator semantics, SSQDA trace computation, and HR-QDA/GQDA distinction against primary sources.
• Added deterministic formula, prediction, invariance, tuning-isolation, solver-certificate, and degeneracy tests. At the earlier Chapter 1--5 gate, the source tarball passed 45 test files, 495 named test blocks, 4,772 expectations, all 170 Rd files, examples and both vignettes with zero errors and warnings; the only NOTE is the existing Windows linked-symbol scan.
• No paper-specific classification simulation grid, error-rate table, reproduction script, benchmark data, or guessed optimizer was added.

Chapter 6 dimension reduction

• Added 16 callable APIs across four modules: classical PCA/CCA/Bartlett testing and robust factor/subspace/number methods; spatial-sign, Kendall and generalized-sign PCA; TPM/Fantope/PMD sparse PCA and CCA; and the primary metric-lasso sscca() kept distinct from the book's whitened-PMD variant.
• Made centers, divisors, ranks, eigengaps, ties, supports, l1 boundaries and ridge operators explicit. KKT, PSD and convergence certificates are returned; invalid fits are never promoted or silently repaired.
• Separated generalized-sign median/raw-MAD and h-order cutoff families and recorded the radial population scope. Preserved the primary SSCCA metric diagonals and BIC program rather than silently whitening them.
• Added 77 named deterministic test blocks with 539 static source expect_* call occurrences. They verify formulas, invariances, solvers and boundaries; no paper Monte Carlo design, tuning result, table, script or data was added.

Chapter 7 clustering

• Added Lloyd K-means, full-covariance Gaussian-mixture EM, sparse K-means and its selector, sparse K-medians, CHIME, IF-PCA, K-spatial medians, SM-SSCM, Sparse-SM and its tau/K selectors, plus SEMC fitting/prediction and Gap-LSE selection: 14 exported APIs and 19 registered compiled kernels.
• Canonicalized CHIME labels after initialization and every update so omega <= 1/2, with coherent mean, responsibility, beta, score, class and history orientation. The full supplied lambda path and exact lasso KKT certificates are preserved without a hidden ridge.
• Verified ordered-pair BCSS as 2 * (TSS - WCSS), the s=1 soft-threshold boundary, full-covariance GMM rank/SPD checks, and explicit tie, start, farthest-empty, cycle and returned-final-state contracts.
• Hardened robust spatial geometry with raw-first close-large differences, overflow-safe scaled accumulation, a Cholesky metric root, and finite subnormal/opposite-DBL_MAX paths.
• Kept selector permutations and IF-PCA null generation local, fixed-seed and caller-RNG preserving. These are intrinsic calibrations, not paper simulations.
• Added semc_fit(), predict.semc_fit(), and semc_select_k_gap() as an independent rewrite against the primary source and pinned flnankai/GEMcluster@10fce04fe690fe274dd5d237cfcd3d5c6a4139f6 MIT release. Primary-paper and software-only controls remain explicit and separate; fixed formula/software fixtures contain no copied author-code expression or paper simulation.
• The fresh unified gate has 176 exports, 3 S3 registrations, 140 Rcpp attributes matched by 140 native registrations, 223 Rd files, and 169 public topics. Its 56 test files contain 688 test_that blocks and 5,781 static expect_* calls; the installed run reported 6,267 passes and zero failures, errors, warnings, or skips. Two generation passes produced an identical 226-file hash manifest and all 226 files matched the release tree.
• The final 87-row ledger records 83 implemented, 2 review-only, and 2 source-blocked method rows. The blocked rows are the Chapter 2 structured- correlation review and Chapter 4 mutual-independence studentization; SEMC is implemented.
• The final 0.1.0 source tarball built from this release state passed standard R CMD check --no-manual with 0 ERROR, 0 WARNING, and 1 NOTE, and remote-enabled R CMD check --as-cran --no-manual with 0 ERROR, 0 WARNING, and 2 NOTEs. Both ran examples, tests, and vignettes, and the as-CRAN check also ran donttest. Both checks share the conservative DLL linked-symbol scan for _exit, abort, and exit; exact R/src direct-call scanning found zero. The additional as-CRAN NOTE identifies a new CRAN submission.
• The public flnankai/HDElliptical repository and issue tracker were bootstrapped before the final tag. The post-bootstrap as-CRAN check kept incoming remote checks enabled, and the Windows, macOS, and Linux Actions jobs passed before the annotated v0.1.0 tag, release, and regenerated release assets were published.
• Added a deterministic installed-package benchmark harness and checked-in CSV/ Markdown report with one representative workflow per chapter. These are machine-local software timings and fingerprints, not paper simulations, size/power experiments, or empirical reproductions.