Projected Hotelling T-squared divide-and-combine confidence region
Mathematical Methods AIMech:ProjectedHotellingT2DivideAndCombineConfidenceRegion DRAFT
Aliases: Half-Cauchy Divide-and-Combine Multivariate Confidence Region, Projected Hotelling's T^2 Half-Cauchy Confidence Region, Divide-and-Combine Subspace Hotelling T-Squared Confidence Set, HCCT Divide-and-Combine Confidence Region
A multivariate statistical inference and meta-analysis method (Liu, Meng, & Pillai 2025) that constructs a guaranteed convex, bounded joint confidence region C_alpha(theta) in R^d for a d-dimensional parameter vector theta by projecting or partitioning study estimators into m lower-dimensional linear subspaces P_j * theta in R^{d_j} (where P_j in R^{d_j x d} has full row rank d_j <= d), calibrating each sub-study's quadratic pivot ||Sigma_hat_j^(-1/2)(xi_hat_j - P_j * theta)||^2 against a finite-sample Hotelling's T^2(d_j, k_j) distribution (or chi^2_{d_j} when k_j = infinity) with CDF F^(j), and inverting the simplex-weighted Half-Cauchy combination statistic T_HCCT(theta) = sum_{j=1}^m w_j tan((pi / 2) * F^(j)(||Sigma_hat_j^(-1/2)(xi_hat_j - P_j * theta)||^2)) <= F_{HC,w}^(-1)(1 - alpha) across arbitrarily dependent sub-studies without estimating or inverting the global d x d covariance matrix.
Identity Notes & Concept Scope
No normalized identity match for Projected Hotelling T-Squared Divide-and-Combine Confidence Region across all 442 pinned AIO records and 569 maintained IDs, names, aliases and scoped synonyms. Fixed-string searches across data, sources, research, history, and reviews returned only cross-references in AIMech:HalfCauchyCombinationTest and predecessor discovery reports. Distinct from AIMech:HalfCauchyCombinationTest, which defines the scalar heavy-tailed p-value combination test T_HC(p; w) = sum_{j=1}^m w_j tan((pi/2)(1 - p_j)) for global null hypothesis testing rather than the multivariate subspace-projection Hotelling T^2(d_j, k_j) test-inversion geometry (Theorem 3.2 convexity under k_j >= d_j + 1, n < d high-dimensional divide-and-combine mean inference in Section 4, network meta-analysis contrast synthesis in Section 5, and Proposition 3.3 adaptive stopping-time conflict downweighting). Distinct from AIMech:CauchyCombinationTest, whose two-sided test inversion diverges to -infinity at each sub-study estimate and fragments into up to m + 1 disconnected confidence regions (Example 1, Section 2.2).
Claims & Grounded Evidence
4 Claims
Mechanism
#hotelling-t2-half-cauchy-composition-convexity-and-compactness
For m sub-studies estimating P_j * theta in R^{d_j} via xi_hat_j and positive-definite covariance estimator Sigma_hat_j with k_j degrees of freedom, calibrating ||Sigma_hat_j^(-1/2)(xi_hat_j - P_j * theta)||^2 via Hotelling's T^2(d_j, k_j) ~ (d_j k_j / (k_j + 1 - d_j)) F(d_j, k_j + 1 - d_j) (or chi^2_{d_j} for k_j = infinity) with CDF F^(j) and composing with the Half-Cauchy quantile function F_HC^(-1)(u) = tan(pi * u / 2) yields a strictly convex objective T_HCCT(theta) on R^d whenever k_j >= d_j + 1 for all 1 <= d_j <= d (Theorem 3.2 and Lemma D.3, proven via curvature dominance T_{F_HC}(u) >= T_{H_{d_j,k_j}}(u) >= T_{H_{d_j}}(u)), whereas inverting the standard two-sided Cauchy combination test produces up to m + 1 disconnected confidence intervals because F_C^(-1)(1 - p_j(theta)) -> -infinity at theta = theta_hat_j (Section 2.2, Example 1); furthermore, C_alpha(theta) is compact (bounded) whenever the row spaces {Row(P_j) : w_j > 0} span R^d (Lemma D.4).
Mechanism
#high-dimensional-small-sample-divide-and-combine-without-global-covariance-inversion
For a single d-dimensional sample X_1, ..., X_n of size n where classical d-dimensional Hotelling's T^2 requires n >= d + 1 and full d x d sample covariance inversion, projecting the sample into m virtual sub-studies of dimension d_j << d via P_j * X_i requires only n >= max_j d_j + 2 (specifically n >= 3 for coordinate or 1D random projections d_j = 1 even when d >> n) and remains valid under rank-deficient global covariance Sigma provided each projected covariance P_j Sigma P_j^T is positive definite (Section 4.1–4.3); simultaneous (1 - alpha) confidence intervals for any linear contrast b^T theta are recovered without Bonferroni multiplicity penalties by minimizing and maximizing b^T theta over the convex set C_alpha(theta) (Section 3.2, lines 394–399).
Mechanism
#inter-study-conflict-lower-bound-and-stopping-time-downweighting-coverage
Because T_HCCT(theta) >= 0 on R^d, severe inter-study conflict can cause min_theta T_HCCT(theta) > F_{HC,w}^(-1)(1 - alpha) and produce an empty confidence region, diagnostic of study heterogeneity or model misspecification and lower-bounded in 1D by min_theta T_HCCT(theta) >= sum_{j=1}^m w_j |theta_hat_j - theta_hat_med| / sigma_hat_j (Section 3.3, Inequality 3.7); for any adaptive sequence of simplex weights w^(0), w^(1), ... that iteratively downweights the study with the largest summand by a factor gamma in (0, 1) until the stopping time tau of the first non-empty region R^(tau), the union confidence region R* = union_{k=0}^tau R^(k) preserves valid coverage P(theta_0 in R*) >= 1 - alpha because R^(0) subset R* (Section 3.3, Proposition 3.3; Appendix D, lines 2142–2146).
Application
#network-meta-analysis-and-exact-harmonic-mean-divide-and-combine-extension
In network meta-analysis comparing d + 1 treatments across m head-to-head or multi-arm trials with design matrix Omega in R^{m x d} whose j-th row omega_j^T encodes the active treatment contrast P_j = omega_j^T, Projected Hotelling T^2 divide-and-combine synthesis combines trial-level summary statistics (xi_hat_j, sigma_hat_j^2) into joint convex confidence regions and simultaneous pairwise treatment contrast intervals without knowing within-trial or cross-trial correlations (Section 5, lines 450–466); replacing F_HC^(-1) with the exact harmonic mean p-value calibration 1 / (1 - F^(j)) similarly yields convex divide-and-combine confidence regions under k_j >= d_j + 1 (Appendix A, Theorem A.3, lines 1677–1680).
⚠️ Curated Limitations & Negative Results
- Convexity of each sub-study's composed Half-Cauchy–Hotelling T^2 objective requires degrees of freedom k_j >= d_j + 1 (so sample size n >= d_j + 2 when k_j = n - 1), and boundedness (compactness) of the combined confidence region C_alpha(theta) requires that the union of row spaces {Row(P_j) : w_j > 0} spans R^d (Theorem 3.2 and Lemma D.4).
- When sub-study estimates exhibit strong heterogeneity or model misspecification, min_{theta in R^d} T_HCCT(theta) can exceed F_{HC,w}^(-1)(1 - alpha), yielding an empty initial confidence region R^(0); resolving emptiness via Proposition 3.3 sequential outlier downweighting requires specifying a weight-decay schedule gamma in (0, 1) and can widen the resulting region toward the remaining studies (Section 3.3, lines 401–427).
- Because the critical value F_{HC,w}^(-1)(1 - alpha) ~ 2 / (pi * alpha) is calibrated for asymptotic tail validity as alpha -> 0+ under arbitrary dependence, coverage at moderate nominal levels (such as 1 - alpha = 0.90 or 0.95) and larger subspace counts m can be slightly conservative or require small-m / finite-alpha adjustments when sub-studies are strongly positively correlated (Section 4.2–4.3).
- All 2,200 lines of Liu, Meng, & Pillai (2025, arXiv:2501.01065v1) were read. Constructed numerical checks in research/discovery/projected-hotelling-t2-divide-and-combine-boundaries.py verify Section 2.2 Example 1 (1 convex HCCT interval vs. 3 disconnected CCT intervals), Theorem 3.2 / Section 4 small-n (d = 6, n = 4 < d + 1) divide-and-combine convexity and simultaneous contrast intervals, and Section 3.3 Inequality (3.7) and Proposition 3.3 stopping-time outlier downweighting.
Artificial Intelligence Ontology (AIO) Grounding
Mapped AIO Classes
| AIO Target Class | Mapping Relation | Mapping Rationale |
|---|---|---|
AIO:DataDredgingBias
Data Dredging Bias
|
related_match | Projected Hotelling T^2 divide-and-combine confidence regions yield valid simultaneous confidence intervals across arbitrary linear contrasts b^T theta and correlated sub-studies without post-hoc selection inflation (AIO:DataDredgingBias). AIO:DataDredgingBias denotes the multiple-testing bias rather than the multivariate confidence-region methodology. |
Evidence Sources & Primary Literature
2 Sources| Source ID | Publication / Document | Type / Access | Reference & Checksum |
|---|---|---|---|
liu-meng-pillai2025
|
Molei Liu, Xiao-Li Meng, Natesh S. Pillai
|
primary paper
full_text
|
ID: https://doi.org/10.48550/arXiv.2501.01065
sha256:2fe912b58a74…
|
fang-chang-park-tseng2023
|
Yusi Fang, Chung Chang, Yongseok Park, George C. Tseng
|
primary paper
full_text
|
ID: https://doi.org/10.5705/ss.202021.0078
sha256:432b18ec119b…
|
Ontology & Grounding Topology
Open in interactive network graph →Curation Provenance
Curator: Antigravity
Method: ai_assisted
Model: Gemini 3
Created: 2026-10-17T10:50:00Z
Updated: 2026-10-17T10:50:00Z
Source File:
data/curated/ProjectedHotellingT2DivideAndCombineConfidenceRegion.yaml
· Schema:
curated_concept.yaml
Explore across X-Mech
Related computational & biological Mechs:
Embedding Map
Network Graph
ProteinTraitsMech (ESM-2 pLM)
DUFMech (Domain pLM)
TraitMech
CultureMech
All 14 X-Mech Repositories