Where Do Two Populations of Persistence Diagrams Differ? Calibrated Local Inference at a Fixed Budget

As an AI researcher, I have meticulously analyzed the provided excerpts from two distinct sections (A and B) of a research paper titled Where Do Two Populations of Persistence Diagrams Differ? Calibrated Local Inference at a Fixed Budget. My task is to synthesize these findings into a single, comprehensive, and highly detailed summary.

Here is the detailed synthesis:

Comprehensive Research Summary: Calibrated Local Inference for Persistence Diagrams at a Fixed Budget

This research investigates advanced statistical methods for performing simultaneous inference on local mean contrasts within populations of persistence diagrams, specifically addressing the challenge of identifying the precise regions on the birth–death plane that drive these differences. The core methodology focuses on deriving localized statistical tests when the total number of available diagrams is fixed.

Core Methodology and Local Contrast Definition (From Section A)

The study moves beyond traditional two-sample tests that only assess global differences. Instead, it defines local mean contrasts as the expected weighted feature mass differences within infinity neighborhoods centered at various points and radii on the birth–death plane. These contrasts are estimated using additive landmark responses derived from methods like PLACE and PALACE.

To ensure rigorous statistical control, a Gaussian multiplier bootstrap is employed to calibrate simultaneous confidence intervals, allowing for unequal group covariances. A crucial finding is that the neighborhoods whose confidence intervals exclude zero form a map that approximates family-wise error control. Furthermore, the study demonstrates that selecting a subset of these original intervals for display preserves their joint coverage guarantee. Geometrically, this leads to a key result: on the simultaneous coverage event, every reported neighborhood lies within twice its radius of the support of the mean-measure difference. A geometric condition also establishes sufficient radius requirements for a displaced feature to yield a non-zero contrast.

Performance and Coverage Guarantees (From Section A)

The study rigorously validates its approach through simulations involving 40 to 120 diagrams per class. The localized region successfully concentrates on rings of fused-ring systems when applied to the MUTAG benchmark, suggesting a structural interpretation of the local contrasts. The achieved simultaneous coverage rates are high, reaching 94%–98% under both strict null and equal means with unequal covariances. A significant theoretical result confirms that localized region selection can enhance power when several nearby changes share the same sign. The paper also establishes bounds for simultaneous coverage and detection probability across different normalization regimes (fixed or estimated).

Comparison with Alternative Inference Methods (From Section B)

The paper contextualizes its findings by comparing the persistence diagram inference methods against established alternatives:

1. Pixelwise Image Inference (F.6): The study compares the proposed band method against an image comparator that applies one test per persistence-image pixel, using a Westfall–Young permutation maximum for family-wise calibration under exchangeability. When testing with m=20 coordinates, the image map shows power between 70–94% at = 0.10 but drops significantly to 15–23% at = 0.20. The band approach shows superior performance in localization error (about 0.04) compared to the image-based landmark inference (about 0.118 for m=63).

2. Lifetime-Changing and Lifetime-Preserving Shifts (F.7): A comparison of six procedures (including the band's global maximum test, MMD with an additive Gaussian kernel, log-rank statistics, landscape MMD, and persistence-image MMD) across different regimes (m per class) shows that all tested procedures maintain null rejection rates between 0.02 and 0.07. The band method demonstrates power at = 0.10 and = 0.20, suggesting robust performance across varying shift types in the persistence regime.

3. Cross-Fitting and Unequal-Covariance Nulls (F.8): This section explores a complex scenario where data is split into two halves, with landmarks/normalization fitted on one half and tested on the other at alpha/2. The reported localization set is the union of both sets, ensuring every diagram supplies inference once. Under these cross-fitting conditions—where mean measures agree but covariances differ—the null False Discovery Rate (FWER) is controlled between 0.014–0.046 across a wide range of coordinate counts (K=50 to 400). The pilot-normalized bootstrap method yields FWER control, contrasting with the Normal-reference Bonferroni approach which is less variable when group sizes are small.

Conclusion and Significance

In summary, this work provides a sophisticated framework for calibrated local inference in persistence diagram analysis. It successfully identifies the specific geometric regions of the birth–death plane responsible for observed differences, achieves high simultaneous coverage (94%–98%), and demonstrates superior performance compared to pixelwise image inference methods when dealing with unequal group covariances and complex null hypotheses arising from cross-fitting procedures. The findings are applicable across various molecular benchmarks (MUTAG, DHFR, COX2, PTC), establishing a powerful tool for pinpointing the source of persistence diagram variation.

1. Inference-based normalization for local contrasts can be adapted to automatically select optimal neighborhood radii by maximizing a pilot score, as described in Choosing a configuration on a pilot (Section 3). This allows AI systems to dynamically adjust their sensitivity based on available data budget and the uncertainty of the underlying feature displacement.

2. The system can perform simultaneous inference for local mean contrasts across birth–death locations and neighborhood radii by using additive landmark responses, which estimates differences in expected weighted feature mass within l∞ neighborhoods at several centers and radii (Abstract). This moves beyond global tests to identify the precise regions of persistence diagrams contributing to observed differences.

3. The AI system can be calibrated using a Gaussian multiplier bootstrap that allows for unequal group covariances, as detailed in the paper, ensuring simultaneous confidence intervals are robust even when groups have different variances (Abstract). This provides more reliable uncertainty quantification than standard tests under complex data structures.

4. The system can employ a selection on the inference sample strategy where it retains coordinates with the largest Zk values, which is shown to give FWER 0.13–0.24 at ∆ = 0 (Table 16). This targeted selection prioritizes features with the strongest evidence of displacement, potentially increasing power for detecting significant changes in sparse data regimes.

5. The AI system can utilize a cross-validated region and radius selection procedure to improve localization, as demonstrated in Appendix D.5, where cross-fitted selection gained power for several nearby changes of the same sign (Appendix D.5). This helps in scenarios with multiple simultaneous localized features by optimizing the search space based on training and validation data splits.

6. The AI system can implement a survival-based comparator using log-rank statistics on lifetime multisets to test for differences between populations, as described in Section C.3 (C implementation details). This is particularly useful when analyzing temporal or sequential data where the persistence diagram is derived from event times, as it ensures Proposition 3 applies to both (Appendix A.8).

7. The system can incorporate a pilot-based selection mechanism to reduce the number of tested coordinates while preserving joint coverage guarantees, following Corollary 3 (Section 3), which compares sufficient detection thresholds accounting for both the loss of inference observations and the reduction in the number of tested coordinates.