acceptodds
Under review as a conference paper at ICLR 2027

LTCC: Localized Testing of Coordinate Contributions to Conditional Distribution Discrepancy

Abstract

We study whether a prespecified response coordinate contributes to the discrepancy between two multivariate conditional distributions at a fixed covariate profile. We define a localized no-contribution hypothesis that requires preservation of the target coordinate's conditional marginal distribution and its marginal-standardized cross-association with the remaining responses. Population-specific conditional probability integral transforms separate association changes from marginal differences in other coordinates. Characteristic kernels yield an equivalent direct-sum reproducing kernel Hilbert space characterization. We derive an exact von Mises expansion accounting for estimated conditional distributions and construct a cross-fitted one-step estimator. We establish a Gaussian limit under fixed alternatives and a weighted- limit under the null for the squared-norm statistic. A consistent multiplier bootstrap yields asymptotic level and consistency against fixed alternatives. Simulations and applications to large language models and image classification illustrate the proposed testing procedure.

Primary area:
interpretability and explainable AI
Keywords:
cross-fitting, debiased estimation, influence function, multiplier bootstrap

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_crvdOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies whether a prespecified response coordinate contributes to the discrepancy between two multivariate conditional distributions at a fixed covariate profile. The localized null hypothesis requires preservation of the target coordinate’s conditional marginal distribution and its marginal-standardized cross-association with the remaining responses. The proposed test statistic is closely related to the conditional MMD studied in Yan et al. (2022). A multiplier bootstrap is used to approximate the critical value, and the authors establish asymptotic validity under the null and consistency under fixed alternatives.

Critical strengths
  1. Rather than only testing whether two conditional distributions differ, the paper investigates which response coordinates contribute to the discrepancy at a particular covariate profile. The hypothesis concerns not only the conditional marginal distribution, but also its marginal-standardized cross-association with the remaining responses.
  2. The paper establishes the asymptotic validity and consistency of the proposed test method.
Critical weaknesses
  1. The main distinction from existing localized conditional two-sample testing is the inclusion of a marginal-standardized dependence component in the null hypothesis. While this extension is mathematically well motivated, its practical importance is not fully established.
  2. Assumption 4.4 requires the conditional distribution estimator to converge in total variation sufficiently fast for the quadratic remainder to be asymptotically negligible.
  3. The theoretical framework assumes fixed covariate and response dimensions.
  4. When multiple covariate profiles or response coordinates are examined simultaneously, no multiple-testing correction is applied.
Questions for authors
  1. The paper would benefit from clearer justification of when dependence-only coordinate contributions are scientifically meaningful. Moreover, Regarding localized conditional distribution testing component, what is the main methodological distinction between the proposed approach and Yan et al. (2022)?
  2. Does the conditional distribution estimator used in the paper satisfy Assumption 4.4? How sensitive is the proposed test to the estimation error?
  3. What is the computational cost of the proposed testing procedure? How should the bandwidth hh be selected in practice? How is it chosen in the experiments?
  4. If the goal is to identify a set of contributing coordinates or covariate profiles, how should multiplicity be handled?
  5. The theoretical analysis focuses on fixed alternatives. It would be useful to provide a local alternative analysis to characterize the detection boundary or asymptotic power. Yan et al. (2022) imposes restrictions relating the kernel smoothness order ν\nu to the covariate dimension. Why are analogous restrictions not required here?
  6. Section 5.2 compares LTCC only with Marginal Local-CMMD. It would strengthen the empirical evaluation to include competing methods in the other simulation and application settings as well.
4 comments
AuthorsOpenReview ↗

Response to Reviewer crvd (Part 1/4)

We appreciate the reviewer’s constructive comments and helpful suggestions. We address each point below.

W1/Q1. Practical Importance and Scientific Interpretation of Dependence-Only Contributions

We appreciate the request to clarify the scientific relevance of dependence-only changes. They are meaningful when the object of interest is a system of responses rather than individual marginal distributions: for example, a fault can preserve the marginal distributions of several sensor measurements while changing their dependence structure, and the association pattern can then identify which measurements participate in the altered system behavior.

For example, in precision medicine, the two populations may represent treatment groups or disease states, with age, sex, and routine clinical measurements as covariates and biomarkers, omics features, or imaging measurements as the multivariate response. Identifying which measurements contribute to the between-group conditional distributional discrepancy can help characterize clinically relevant heterogeneity and guide further investigation. In credit-risk analysis, the populations may represent defaulters and nondefaulters, with age and credit limit as covariates and repayment delays, account balances, and payment histories as the response vector. Identifying contributing behavioral measurements among comparable customers can help characterize patterns associated with default risk. In both settings, a response coordinate may have an unchanged conditional marginal distribution while its marginal-standardized cross-association with the remaining responses changes. We have also added a discussion of the practical relevance of dependence-only contributions to the Introduction of the revised manuscript.

LTCC addresses this problem through three methodological contributions. First, we formulate a localized no-contribution hypothesis that characterizes whether a prespecified response coordinate contributes to a conditional joint-distribution discrepancy through its conditional marginal distribution or its marginal-standardized cross-association. Second, we introduce population-specific conditional PITs to remove coordinatewise marginal effects before comparing cross-associations, preventing marginal changes in non-target coordinates from being incorrectly attributed to the target coordinate. Third, we establish an equivalent direct-sum RKHS characterization and develop a cross-fitted one-step estimator with influence-function correction, together with asymptotic theory and multiplier-bootstrap calibration under the stated assumptions.

Section 5.2 isolates this setting: every coordinate has the same N(0,1)N(0,1) conditional marginal distribution in both populations, while the dependence between Y1Y_1 and Y−1Y_{-1} changes. LTCC has increasing empirical power as the dependence difference grows, whereas Marginal Local-CMMD remains close to its nominal significance level. The formal PIT-based construction and Gaussian illustrations are given in our response to Reviewer HCnF, Q1; here the point is the scientific interpretation of the resulting coordinate contribution. The ImageNet-9 application then illustrates the same coordinate-localization target in structured model outputs.

Q1. LTCC versus Yan et al. (2022)

We appreciate the comparison with Yan et al. (2022). Although both methods use localized conditional distribution testing, LTCC differs in three aspects:

  • Different inferential targets. Yan et al. (2022) test whether conditional response distributions differ at a given covariate profile. LTCC further identifies whether a prespecified coordinate contributes through its conditional marginal distribution or its marginal-standardized cross-association. Applying Yan et al.'s procedure separately to scalar coordinates can miss dependence-only contributions.

  • Marginal-standardized cross-association. We introduce a new coordinate-level measure based on population-specific conditional PITs. By transforming all response coordinates to uniform marginal distributions, this construction prevents marginal changes in non-target coordinates from being incorrectly attributed to association changes involving the target coordinate.

  • Inferential construction. We derive an RKHS-valued von Mises expansion and develop a cross-fitted one-step estimator with influence-function correction and multiplier-bootstrap inference. This differs from the conditional U-statistic and pooled local bootstrap used in our scalar Local-CMMD implementation.

AuthorsOpenReview ↗

Response to Reviewer crvd (Part 2/4)

W2. Convergence Conditions for First-Stage Estimation in Total Variation

We clarify the role of the high-level condition on the conditional-distribution estimator and provide finite-sample sensitivity evidence. Assumption 4.4 requires the cross-fitted conditional-distribution estimators to have error Op(rn)O_p(r_n) in total variation at x0x_0; Assumption 4.5(iii) adds an−1rn2→0a_n^{-1}r_n^2\to0. Together, these conditions control the quadratic remainder in Theorem 3.1 and, with cross-fitting and the remaining conditions in Assumption 4.5, the perturbation of the estimated influence functions in the expansion and bootstrap. The detailed double-kernel construction, regularity conditions, and convergence-rate calculation showing attainability are given in our response to Reviewer UCJ7, W2. Figure 6 provides complementary finite-sample evidence for the estimators used in the experiments.

Q2. Sensitivity to First-Stage Estimation Error

Figure 6 complements Assumption 4.4 with a finite-sample sensitivity study. Under the coordinate null hypothesis, LTCC has empirical rejection rates close to the nominal significance level with the oracle and flexible estimators, whereas the misspecified linear-Gaussian estimator leads to size distortion. Under the shape alternatives, the flexible estimators retain increasing empirical power. The detailed rate argument and KDE attainability example are given in our response to Reviewer UCJ7, W2; Figure 6 reports the estimator-sensitivity results in Appendix D.4.

W3. Fixed-Dimensional Theoretical Scope

We thank the reviewer for pointing out this scope limitation. Our theoretical framework establishes pointwise inference for fixed covariate and response dimensions dd and pp under Assumptions 4.1–4.5. Extending LTCC to diverging covariate or response dimensions is important, especially for high-dimensional applications, and would require additional control of the conditional-distribution estimation error, localization, and influence-function terms. We consider this a natural direction for future research. The concluding discussion also clarifies the scope of our current theoretical results and notes the potential extension to diverging dimensions.

Q3. Computational Cost and Bandwidth Selection

We thank the reviewer for asking about computational cost and bandwidth selection. In the revised version, we have added a detailed computational-complexity analysis to Appendix D.4.

For one prespecified response coordinate jj and covariate profile x0x_0, let n=n1+n2n=n_1+n_2 denote the total sample size, MM the number of conditional Monte Carlo draws, and BB the number of multiplier-bootstrap repetitions. Let TfirstT_{\mathrm{first}} denote the estimator-dependent cost of fitting the first-stage conditional distributions and generating the corresponding Monte Carlo draws.

For the Gaussian-kernel implementation, the main computational cost arises from constructing the influence-function Gram matrix. Direct pairwise evaluation gives a conservative upper bound of O(p2n2M2)O(p^2n^2M^2). Once constructed, the Gram matrix is reused across bootstrap repetitions, each requiring O(n2)O(n^2) operations. The fitted conditional distributions and Monte Carlo draws are also reused across bootstrap repetitions.

Under the fixed-dimensional assumption (Assumption 4.1), the overall computational complexity therefore simplifies to TLTCC=Tfirst+O(n2(M2+B)).T_{\mathrm{LTCC}}=T_{\mathrm{first}}+O\left(n^2(M^2+B)\right).

For practical bandwidth selection, we use a sample-size-dependent bandwidth of order hℓ≍nℓ−1/4h_\ell\asymp n_\ell^{-1/4}, which satisfies the stated fixed-dimensional bandwidth conditions for the dimensions and kernel order considered. The specific bandwidth choices for the synthetic experiments and ImageNet-9 application are detailed in Appendices D.1 and D.5, respectively. Appendix D.4 further examines sensitivity to the localization and association-kernel bandwidths, with empirical size and power reported in Figure 7.

AuthorsOpenReview ↗

Response to Reviewer crvd (Part 3/4)

W4/Q4. Multiple Testing Across Coordinates and Simultaneous Inference Across Profiles

We thank the reviewer for raising this issue. Our framework establishes pointwise inference for a prespecified coordinate jj and profile x0x_0. Multiple coordinates and profiles lead to two inferential problems:

(1) Multiple testing across response coordinates.

For a fixed profile x0x_0, the family consists of the pp coordinate null hypotheses. Bonferroni and Holm adjustments can be directly applied to the pointwise p-values to control the family-wise error rate (FWER), without requiring independence among the tests. The Benjamini–Hochberg (BH) procedure can provide false discovery rate (FDR) control under suitable dependence conditions, such as independence or positive regression dependence on a subset (PRDS). However, the dependence among LTCC coordinate-wise p-values is nontrivial because the tests share observations and involve overlapping association components. Therefore, the standard BH guarantee cannot be directly invoked without further characterizing this dependence structure.

(2) Simultaneous inference across a finite set of covariate profiles.

In the revised version, we have established simultaneous inference across a finite set of covariate profiles and provided the corresponding theoretical guarantees in Appendix H.

For a fixed response coordinate jj consider a prespecified finite collection of distinct covariate profiles X0=x1,…,xK\mathcal X_0=\\{x_1,\ldots,x_K\\}, with fixed KK. Define the joint null hypothesis H0,j(X0):   Qj(x)=0,∀x∈X0.H_{0,j}(\mathcal X_0):~~~Q_j(x)=0,\forall x\in\mathcal X_0.

Under Assumptions 4.1–4.5 imposed at each profile, with the first-stage total-variation convergence condition holding jointly over this finite collection, the existing influence-function expansion extends to the product Hilbert space HjK\mathcal H_j^K. In particular, writing γ^j,n=(γ^j(x1),…,γ^j(xK))\widehat{\boldsymbol\gamma}_{j,n}=\big(\widehat\gamma_j(x_1),\ldots,\widehat\gamma_j(x_K)\big), γj=(γj(x1),…,γj(xK))\boldsymbol\gamma_j=\big(\gamma_j(x_1),\ldots,\gamma_j(x_K)\big), we obtain an−1(γ^j,n−γj)⇝Gjin HjK,a_n^{-1}\big(\widehat{\boldsymbol\gamma}_{j,n}-\boldsymbol\gamma_j\big)\rightsquigarrow\boldsymbol G_j\quad\text{in }\mathcal H_j^K, where Gj\boldsymbol G_j is a joint Gaussian random element with covariance operator Vj=diag⁡Vj(x1),…,Vj(xK)\mathbb V_j=\operatorname{diag}\\{V_j(x_1),\ldots,V_j(x_K)\\}. This block-diagonal structure follows from the compact support of the localization kernel and hℓ→0h_\ell\to0: the local neighborhoods of distinct fixed profiles eventually do not overlap, so their leading influence-function terms have asymptotically zero cross-covariances.

We can then construct a joint multiplier bootstrap by generating observation-level Gaussian multipliers independently within each population and reusing the same multiplier for each observation across all profiles. Specifically, for each xkx_k, we compute the bootstrap process using the existing construction in Section 4.2, replacing x0x_0 by xkx_k. Denoting the resulting joint process by G^j,n∗=(G^j,n∗(x1),…,G^j,n∗(xK))\widehat{\boldsymbol{ \mathbb G}}_{j,n}^*=\big(\widehat{\mathbb G}_{j,n}^*(x_1),\ldots,\widehat{\mathbb G}_{j,n}^*(x_K)\big), the corresponding joint bootstrap consistency result is dBL(L∗(G^j,n∗),L(Gj))→p0.d_{\mathrm{BL}}\left(\mathcal L^*(\widehat{\boldsymbol{\mathbb G}}_{j,n}^*),\mathcal L(\boldsymbol G_j)\right)\xrightarrow{p}0.

Based on this result, define the maximum bootstrap statistic Tj,n∗,max⁡=max⁡1≤k≤K∥G^j,n∗(xk)∥Hj2T_{j,n}^{*,\max}=\max_{1\leq k\leq K}\left\|\widehat{\mathbb G}_{j,n}^*(x_k)\right\|_{\mathcal H_j}^{2}, and let c1−α,n∗,max⁡c_{1-\alpha,n}^{*,\max} be its conditional (1−α)(1-\alpha)-quantile. Under the corresponding nondegeneracy and continuity conditions, we obtain sup⁡t∈R∣P∗(Tj,n∗,max⁡≤t)−P(max⁡1≤k≤K∥Gj(xk)∥Hj2≤t)∣→p0.\sup_{t\in\mathbb R}\left|P^*(T_{j,n}^{*,\max}\leq t)-P\left(\max_{1\leq k\leq K}\|G_j(x_k)\|_{\mathcal H_j}^{2}\leq t\right)\right|\xrightarrow{p}0. Consequently, simultaneous inference across the prespecified profiles can be conducted using the common critical value: Reject H0,j(X0)ifmax⁡1≤k≤Kan−2Q^j(xk)>c1−α,n∗,max⁡.\text{Reject }H_{0,j}(\mathcal X_0)\quad\text{if}\max_{1\leq k\leq K}a_n^{-2}\widehat Q_j(x_k)>c_{1-\alpha,n}^{*,\max}. The resulting max-bootstrap procedure provides asymptotic type I error control at level α\alpha for the joint null hypothesis H0,j(X0)H_{0,j}(\mathcal X_0).

AuthorsOpenReview ↗

Response to Reviewer crvd (Part 4/4)

Q5. Local Alternatives and Kernel-Order Conditions

We thank the reviewer for this helpful suggestion.

(1) Local alternatives.

In the revised version, we have included new theoretical results in Appendix I on the asymptotic properties of our proposed test under the following local alternatives.

Let $ a_n^2=(n_1h_1^d)^{-1}+(n_2h_2^d)^{-1}. $ Consider a local alternative sequence $ \gamma_{j,n}(x_0) $ satisfying

\qquad \Delta_j\in\mathcal H_j.$$ This formulation covers local changes in the conditional marginal distribution, the marginal-standardized cross-association, or both. In Appendix I, we establish uniform versions of the asymptotic linear representation and show that $$a_n^{-1}\widehat\gamma_{j,n}(x_0) \Rightarrow \mathbb G_{j,0}(x_0)+\Delta_j,$$ where $\mathbb G_{j,0}(x_0)$ is Gaussian with mean zero and covariance $V_{j,0}(x_0)$. Hence $$a_n^{-2}\widehat Q_{j,n}(x_0)\Rightarrow\left||\mathbb G_{j,0}(x_0)+\Delta_j\right||_{\mathcal H_j}^{2}.$$ The bootstrap critical value converges to the corresponding null quantile $c^{(0)}_{j,1-\alpha}(x_0)$, and the local asymptotic power is $$\beta_j(\Delta_j)=\mathbb{P}\left\\{\left\||\mathbb{G}_{j,0}(x\\_0)+\Delta_j\right\||_{\mathcal{H}_j}^{2}>c^{(0)}_{j,1-\alpha}(x_0)\right\\}.$$ More generally, write $\Delta_{j,n}:=a_n^{-1}\gamma_{j,n}(x_0)$. The resulting asymptotic power regimes are: - If $\Delta_{j,n}\to0$, equivalently $\|\gamma_{j,n}(x_0)\|=o(a_n)$, the asymptotic power converges to $\alpha$. - If $\Delta_{j,n}\to\Delta_j\neq0$, as in the local-alternative formulation above, the asymptotic power converges to the non-trivial value $\beta_j(\Delta_j)$ defined above. - If $\|\Delta_{j,n}\|\to\infty$, equivalently $\|\gamma_{j,n}(x_0)\|/a_n\to\infty$, then, provided the estimation error remains uniformly $O_p(a_n)$ and the bootstrap critical values remain bounded in probability, the asymptotic power converges to one. Thus, $a_n$ is the local detection scale for the RKHS contrast, while $a_n^2$ is the corresponding scale for $Q_j(x_0)$. **(2) Kernel-order conditions.** Yan et al. (2022) require $\nu>d/2$ for their global statistic, whereas their local procedure has no such additional restriction. LTCC assumes a bounded, compactly supported localization kernel of order $\nu\ge2$. Assumption 4.3 requires $\nu$-times continuous differentiability with bounded derivatives for the covariate density and for the conditional-distribution map in total variation. Assumption 4.5(ii) imposes \$ n_\ell^{1/2}h_\ell^{\nu+d/2}\to0 \$ together with stable local effective sample sizes. These conditions control the pointwise smoothing bias and variance; hence no separate $\nu>d/2$ condition is required for LTCC’s local procedure. ### Q6. Targeted Baseline Comparisons In addition to the comparison with Marginal Local-CMMD in Section 5.2, we also compare LTCC with Zipper-LOCO in Appendix D.2. Section 5.2 shows that LTCC detects dependence-only contributions when conditional marginal distributions are unchanged, whereas Marginal Local-CMMD remains close to its nominal significance level. Appendix D.2 shows that LTCC continues to identify a distributionally contributing coordinate as redundancy increases, whereas Zipper-LOCO's empirical power decreases as incremental predictive importance diminishes. If any additional baseline methods are considered relevant for comparison, we would greatly appreciate the suggestions and would be happy to include them in additional simulations. --- We thank the reviewer again for the constructive suggestions, which helped us clarify the practical interpretation of LTCC and its relation to existing methods.
Reviewer_UCJ7OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies whether a prespecified response coordinate contributes to a discrepancy between two populations' conditional joint distributions at a fixed covariate profile, by formulating a localized no-contribution hypothesis. Using conditional probability integral transforms (PITs), the authors characterize changes in the target coordinate's conditional marginal distribution and its marginal-standardized cross-association with the remaining coordinates, and construct a testing target through kernel mean embeddings. Combining local kernel weighting, cross-fitting, and an influence-function one-step correction, they establish asymptotic distributions and a multiplier bootstrap test. Simulations and model-output audits illustrate the method's ability to detect both marginal changes and pure dependence changes.

Critical strengths

The statistical question is well posed. The coordinate-level target goes beyond marginal comparison and is explicitly distinguished from predictive variable importance.

The inferential construction is fairly complete: it accounts for PIT estimation and handles the degeneracy of the statistic under the null.

The experiments examine size, power, and estimation error, and the appendix provides analyses of first-stage sensitivity, bandwidth sensitivity, and the benefit of the one-step correction.

Critical weaknesses
  1. A gap remains between theory and computation. As clarified in Appendix D.4, the asymptotic results in Section 4 concern an estimator with expectations under the fitted conditional distributions evaluated exactly. The implementation instead approximates these expectations and marginal CDFs using finitely many Monte Carlo draws from the fitted conditional distributions, while the bootstrap conditions on these draws. The effect of this additional approximation on test calibration is not fully addressed.

  2. The theory requires sufficiently fast convergence of conditional joint-distribution estimators in total variation; using a flexible generative model does not automatically establish these conditions. The main experiments use one-dimensional localization, providing limited evidence for settings in which several covariates jointly affect the conditional distribution.

  3. The reported pointwise tests do not provide simultaneous error control across coordinates and covariate profiles, limiting their interpretation as a joint set of coordinate discoveries. Some baselines also test different null hypotheses, so power differences should not be interpreted as general superiority. Finally, the LLM experiment is a semi-synthetic audit based on frozen model outputs and a constructed routing mechanism, leaving its external validity to be further established.

Questions for authors
  1. How does the number MM of Monte Carlo draws from the fitted conditional distributions affect the null limiting distribution and bootstrap validity? Can the authors provide sufficient growth conditions linking MM, the observation sample size, and the localization bandwidth, together with sensitivity experiments?

  2. For the conditional-distribution estimators actually used, what concrete assumptions ensure the required convergence rates? How do size and power change when several covariates jointly affect the conditional distribution?

  3. What is the intended family of hypotheses in the ImageNet-9 analysis? If all coordinate--profile tests are treated as one family, which conclusions remain after Bonferroni or BH correction?

  4. Can the authors clarify whether the baseline comparisons demonstrate differences between testing targets or statistical-efficiency gains for a common target?

  5. Building on the related-work discussion, can the authors formally compare the proposed localized no-contribution hypothesis with the hypotheses tested by feature-shift detection methods such as Kulinski et al. (2020), and provide concrete examples in which these hypotheses yield different conclusions? Which steps in the inferential development require new arguments beyond existing kernel-testing and one-step estimation results?

6 comments
AuthorsOpenReview ↗

Response to Reviewer UCJ7 (Part 1/4)

We appreciate the reviewer’s constructive comments and helpful suggestions. We address each point below.

W1. Monte Carlo Approximation in the Implemented Estimator

In the revised version, we have added a finite-MM Monte Carlo approximation-error analysis in Appendix F and a corresponding sensitivity study in Appendix G.1.

Appendix F quantifies the Monte Carlo approximation error in the implemented estimator. Conditional on the fitted conditional distributions, Lemma F.1 (equations (55)–(57)) gives an error of order OPMC(M−1/2)O_{\mathbb P}^{\rm MC}(M^{-1/2}) for the plug-in contrast, the estimated influence functions, and the aggregated one-step estimator. Thus, Appendix F explicitly relates the estimator with exactly evaluated fitted-law expectations to the finite-MM implementation.

Q1. Growth of MM and Finite-Sample Stability

The growth condition concerns Monte Carlo error relative to the sampling scale, while the additional sensitivity analysis assesses the numerical implementation for finite numbers of conditional draws.

  • Asymptotic requirement. Writing an2=(n1h1d)−1+(n2h2d)−1a_n^2=(n_1h_1^d)^{-1}+(n_2h_2^d)^{-1}, condition (63), Man2→∞M a_n^2\to\infty (equivalently M−1/2=o(an)M^{-1/2}=o(a_n)), makes this numerical error negligible at the sampling scale. When the two local effective sample sizes are comparable, this is equivalent up to constants to M/(nℓhℓd)→∞M/(n_\ell h_\ell^d)\to\infty. Corollary F.1 then preserves the Gaussian limiting distribution under fixed alternatives, the weighted-χ2\chi^2 limiting null distribution, and the asymptotic level of the multiplier-bootstrap test under the stated assumptions and the nondegeneracy condition tr⁡(Vj(x0))>0\operatorname{tr}(V_j(x_0))>0.

  • Bootstrap coupling. For the conditional multiplier-bootstrap process, coupling the Monte Carlo and exactly evaluated versions using the same bootstrap multipliers only requires M→∞M\to\infty. The stronger condition Man2→∞M a_n^2\to\infty is needed to make the Monte Carlo error in the observed statistic negligible relative to the sampling scale and hence to preserve calibration under the null hypothesis.

In the numerical implementation, MM is treated as a finite computational tuning parameter; Appendix F establishes a sufficient asymptotic growth condition for MM.

To assess finite-sample stability, Appendix G.1 reports an additional sensitivity study in a Gaussian association-switching design with d=1d=1 and p=5p=5. We test coordinate j=1j=1 at x0=0x_0=0, use n1=n2∈{500,1000}n_1=n_2\in\{500,1000\}, and consider M∈{64,128,256,512,1024,2048}M\in\{64,128,256,512,1024,2048\}. The conditional distributions are fitted with two-fold cross-fitted RealNVP models; we use σF=1\sigma_F=1, σC=0.6\sigma_C=0.6, hℓ=nℓ−1/4h_\ell=n_\ell^{-1/4}, and B=499B=499 multiplier-bootstrap draws. Each configuration uses 200 Monte Carlo replications. Within each replication, all values of MM share the same data, fitted conditional distributions, bandwidths, and bootstrap multipliers, and smaller budgets use nested prefixes of a common pool of 2048 conditional draws. The exact design is given in Appendix G.1, and the results are reported in Table 5:

MMSize (nℓ=500n_\ell=500)Size (nℓ=1000n_\ell=1000)Power (nℓ=500n_\ell=500)Power (nℓ=1000n_\ell=1000)
640.0600.0300.3100.775
1280.0550.0350.3150.825
2560.0500.0350.3100.835
5120.0500.0350.3200.840
10240.0600.0350.3300.835
20480.0500.0350.3200.855

Across the six budgets, empirical size ranges from 0.030 to 0.060. Power ranges from 0.310 to 0.330 for nℓ=500n_\ell=500 and from 0.775 to 0.855 for nℓ=1000n_\ell=1000. For nℓ=1000n_\ell=1000, power is 0.775 at M=64M=64 and ranges from 0.825 to 0.855 for M≥128M\ge128; the empirical size and power at M=512M=512 and M=2048M=2048 differ by at most 0.015 across the four settings. These results provide finite-sample evidence of near-nominal size and stable rejection rates for moderate and large conditional-draw budgets. They complement the sufficient asymptotic condition Man2→∞M a_n^2\to\infty in Appendix F.

AuthorsOpenReview ↗

Response to Reviewer UCJ7 (Part 2/4)

W2/Q2. First-Stage Convergence and Performance with Multiple Covariates

(1) First-stage convergence conditions

Assumption 4.4 is stated as a high-level condition on the fitted conditional distributions, while Figure 6 provides estimator-specific finite-sample evidence for the RealNVP implementation. It requires the cross-fitted conditional-distribution estimators to have error Op(rn)O_p(r_n) in total variation at x0x_0, uniformly over populations and cross-fitting folds. Assumption 4.5(iii) adds an−1rn2→0a_n^{-1}r_n^2\to0. Theorem 3.1 and Appendix B show that the association component of the RKHS-valued von Mises remainder is quadratic in these errors. These bounds, cross-fitting, and the remaining conditions in Assumption 4.5 are used to justify the one-step expansion and inference based on the estimated influence functions.

A concrete example shows how the condition can be attained by a nonparametric estimator. Consider the double-kernel conditional density estimator

\frac{\sum_{i=1}^n K_{b_n}(X_i-x_0)L_{c_n}(Y_i-y)} {\sum_{i=1}^n K_{b_n}(X_i-x_0)},$$ where $K_{b_n}$ and $L_{c_n}$ are product kernels for the $d$-dimensional covariate and $p$-dimensional response, respectively. Under twice continuously differentiable joint and marginal densities near $x_0$, $f_X(x_0)>0$, suitable support and boundedness conditions, second-order kernels, and compatible bandwidths, a representative integrated-$L_1$ convergence rate is $$\left\|\widehat f_{Y\mid X}(\cdot\mid x_0)-f_{Y\mid X}(\cdot\mid x_0)\right\|_{L_1} =O_p\left(b_n^2+c_n^2+\frac{1}{\sqrt{n b_n^d c_n^p}}\right).$$ Since the total variation distance between distributions with densities is one half of the $L_1$ distance between their densities, choosing $b_n\asymp c_n\asymp n^{-1/(4+d+p)}$ gives the representative convergence rate $r_n=O_p(n^{-2/(4+d+p)})$. With bandwidths additionally satisfying $a_n^{-1}r_n^2\to0$, this estimator meets the first-stage requirement. Related conditional-density and multivariate conditional-distribution results include Hall, Racine, and Li (2004) and Racine, Li, and Zhu (2004). This example demonstrates one possible route to attainability of the high-level condition under explicit smoothness assumptions. Figure 6 in Appendix D.4 provides complementary finite-sample sensitivity evidence for the RealNVP implementation, while Assumptions 4.4–4.5 state the corresponding theoretical conditions in Section 4. Hall, P., Racine, J., & Li, Q. (2004). *Cross-Validation and the Estimation of Conditional Probability Densities*. Journal of the American Statistical Association, 99(468), 1015–1026. Racine, J., Li, Q., & Zhu, X. (2004). *Kernel Estimation of Multivariate Conditional Distributions*. Annals of Economics and Finance, 5(2), 211–235. (2) Empirical Size and Power with Two Conditioning Covariates The additional experiment in Appendix G.2 evaluates empirical size and power when two covariates jointly affect the conditional distribution. It remains within the fixed-dimensional theoretical regime. We use $d=2$, $p=5$, $n_1=n_2=4000$, and a conditional Gaussian model with covariate-dependent means and scales and a fixed response correlation matrix. The alternative adds a localized mean shift to $Y_1$, $\Delta(x)=1.2\,b_{0.7}(x_1)b_{0.7}(x_2)$, so that the signal region is $S=(-0.7,0.7)^2$; the conditional copula is unchanged, and coordinates $Y_2,\ldots,Y_5$ satisfy the pointwise null throughout. We test $Y_1$ on the $5\times5$ grid $\{-0.9,-0.4,0,0.4,0.9\}^2$, and test the inactive coordinates at $x_0=(0,0)^\top$. The implementation uses two-fold cross-fitting, conditional RealNVP, $\sigma_F=1$, $\sigma_C=0.6$, product Epanechnikov localization with $h_\ell=1.2n_\ell^{-1/4}$, $M=1024$ conditional draws, and $B=499$ multiplier-bootstrap draws. Each setting uses 200 Monte Carlo replications. Figure 8 reports empirical size averaging 0.046 across the 25 profiles under the exact null, with range 0.015–0.075. Under the localized alternative, power averages 0.799 across the nine profiles inside $S$ and reaches 0.995 at $x_0=(0,0)^\top$; the 16 profiles outside $S$ have average rejection rate 0.048. At the center, the empirical rejection rates for the inactive coordinates $Y_2,Y_3,Y_4,Y_5$ are 0.055, 0.085, 0.045, and 0.065. These results provide pointwise empirical evidence for localization when multiple covariates jointly affect the conditional law.
AuthorsOpenReview ↗

Response to Reviewer UCJ7 (Part 3/4)

W3. LLM-Based Conditional Response Auditing

The large language model (LLM) experiment demonstrates how LTCC can audit coordinate-level changes in structured model outputs. We use precomputed responses from Qwen2.5-7B-Instruct on MMLU-Pro, covering five subjects—Math, Physics, Chemistry, Engineering, and Biology. The conditioning variable XX represents a question-difficulty profile constructed from reference-model margins, and the response vector contains one score for each subject.

The experiment introduces a controlled advice-routing mechanism to examine how coordinate contributions vary across question profiles. Group 1 always uses neutral responses, whereas Group 2 applies auxiliary advice to the Math and Physics coordinates with an activation probability depending on both the routing intensity ρ\rho and the conditioning profile XX. We evaluate the Physics coordinate at x0∈0.25,0.75x_0\in\\{0.25,0.75\\}.

The results show two patterns:

  • Signal localization. Empirical power increases with routing intensity at both profiles, reaching 0.28 at x0=0.25x_0=0.25 and 0.91 at x0=0.75x_0=0.75 when ρ=0.30\rho=0.30.
  • Profile-dependent contributions. The substantially higher empirical power at x0=0.75x_0=0.75 reflects the higher probability of receiving auxiliary advice at this profile. This demonstrates that LTCC can distinguish how the same intervention affects conditional response distributions at different question-difficulty levels.

The frozen response banks retain actual model-generated outputs, while the controlled routing mechanism allows us to examine their coordinate-level effects across question-difficulty profiles. This provides a controlled model-output auditing example that complements the synthetic simulations and the ImageNet-9 application.

Q3. Hypothesis Families in the ImageNet-9 Analysis

At each prespecified foreground-coverage profile, the ImageNet-9 family contains nine coordinate hypotheses. We consider both the three separate nine-test families and the pooled family of 27 coordinate–profile hypotheses. Bonferroni and Benjamini–Hochberg (BH) adjustment at level 0.05 yield no rejections under either family definition. The unadjusted results show a profile-dependent descriptive pattern: Dog, Carnivore, Insect, and Fish have pp-values below 0.05 at low foreground coverage, Primate has p<0.05p<0.05 at middle coverage, and none of the nine coordinates has p<0.05p<0.05 at high coverage. Thus, after multiplicity adjustment, none of these coordinate–profile hypotheses can be reported as a simultaneous discovery. The distinction between pointwise coordinate testing and simultaneous inference across coordinates and covariate profiles, together with the corresponding multiplicity procedures, is further discussed in our response to Reviewer crvd, W4/Q4.

Reviewer_HCnFOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
1
Soundness
3
Read review
Paper summary

The authors study the problem of understanding whether a coordinate j contributes to the discrepancy in the conditional distribution of a multivariate outcome Y=(Y1, ..., Ym) between two environments, conditional on some context X=x0. Was j the coordinate that was responsible for the difference in the two distributions? And if so, was it because its marginal changed or because its association with the other variables changed.

Critical strengths

The problem of localizing the discrepancy, to a point x0 and to a coordinate j is sufficiently motivated. The authors provide formally valid inference results for the hypothesis test they propose.

Critical weaknesses

The particular methodology was not sufficiently motivated. The techniques used are standard conditional kernel mean discrepancy tests for checking whether two distributions are different. The authors simply applied these to two distributions, the marginal of Y_j | X=x0 and to the "cross association measure" that they define, that captures conditional association of Y_j with other variables, in a manner that is "marginal-invariant".

Questions for authors
  1. Why is Cj the right measure of association change? It was very hard to understand, not motivated at all and simply stated as a definition. This is a key part of the formulation and was completely ignored and flew over. Was that formulation from prior work? How does prior work motivate it?
  2. Why do we care to explicitly separate the contribution of j in the discrepancy into these two parts? Why can't we just look at the distribution distance between the distribution of Y | X, and the distribution distance of Y_{-j} | X. The difference in these two distance metrics, is a measure of attribution to coordinate j. That seems to me a much simpler attribution, i.e. looking at d(Y^{(1)}, Y^{(2)} | X=x0) - d(Y_{-j}^{(1)}, Y_{-j}^{(2)} | X=x0), for any distribution distance metric d. This seems to me the extra "distribution distance" created by considering coordinate j.
2 comments
AuthorsOpenReview ↗

Response to Reviewer HCnF (Part 1/2)

We appreciate the reviewer’s constructive comments and helpful suggestions. We address each point below.

W. Coordinate-Level Estimand and Inferential Contribution

Our goal is to determine whether a prespecified response coordinate jj contributes to the discrepancy between two conditional joint distributions at a fixed covariate profile X=x0\boldsymbol{X}=\boldsymbol{x}_0. Such a contribution can arise in two distinct ways: (i) the conditional marginal distribution of YjY_j changes across the two populations, or (ii) the conditional association between YjY_j and the remaining coordinates Y−jY_{-j} changes. Accordingly, our localized no-contribution null hypothesis requires both the conditional marginal distribution and the relevant conditional association to be unchanged. If either component differs, coordinate jj contributes to the conditional distributional discrepancy.

The marginal component is directly characterized by equality of the conditional marginal laws, Fj(1)(⋅∣x0)=Fj(2)(⋅∣x0)F_j^{(1)}(\cdot\mid \boldsymbol{x}_0)=F_j^{(2)}(\cdot\mid \boldsymbol{x}_0). For the association component, a natural raw-scale contrast is the signed measure F(ℓ)−Fj(ℓ)⊗F−j(ℓ)F^{(\ell)}-F_j^{(\ell)}\otimes F_{-j}^{(\ell)}, which measures the departure of the conditional joint distribution from independence between YjY_j and Y−jY_{-j}. However, this contrast is affected by changes in the marginal distributions.

To illustrate this issue, Appendix A considers a three-dimensional Gaussian setting in which Y1Y_1 and Y2Y_2 may be correlated, while Y3Y_3 is independent of (Y1,Y2)(Y_1,Y_2). The two populations may differ in the correlation between Y1Y_1 and Y2Y_2 and in the marginal distribution of Y3Y_3. When the correlations satisfy ρ1=ρ2=ρ≠0\rho_1=\rho_2=\rho\neq0 and the marginal shift satisfies δ≠0\delta\neq0, only the marginal distribution of Y3Y_3 differs, while the dependence structures remain identical. Nevertheless, Proposition A.2 shows that the raw-scale association measures for coordinates 1 and 2 differ across populations. Directly comparing these measures would therefore incorrectly attribute the marginal change in Y3Y_3 to association changes involving Y1Y_1 and Y2Y_2.

This motivates applying population-specific conditional probability integral transforms (PITs) to every response coordinate before comparing associations. Let Fr(ℓ)F_r^{(\ell)} denote the conditional marginal CDF of Yr(ℓ)Y_r^{(\ell)} at x0\boldsymbol{x}_0, and define Ur(ℓ)=Fr(ℓ)(Yr(ℓ))U_r^{(\ell)}=F_r^{(\ell)}(Y_r^{(\ell)}), r∈[p]r\in[p], and write U(ℓ)=(U1(ℓ),…,Up(ℓ))⊤\boldsymbol{U}^{(\ell)}=(U_1^{(\ell)},\ldots,U_p^{(\ell)})^\top. Each transformed marginal is uniform, so marginal-scale differences are removed before association is compared. Let G(ℓ)G^{(\ell)} and G−j(ℓ)G_{-j}^{(\ell)} denote the conditional distributions of U(ℓ)\boldsymbol{U}^{(\ell)} and U−j(ℓ)\boldsymbol{U}_{-j}^{(\ell)} at x0\boldsymbol{x}_0, and define Cj(ℓ)=G(ℓ)−U⊗G−j(ℓ)C_j^{(\ell)}=G^{(\ell)}-\mathcal{U}\otimes G_{-j}^{(\ell)}, where U\mathcal{U} is the uniform distribution on [0,1][0,1]. The product measure preserves the distribution of the remaining transformed coordinates while removing their association with the target coordinate.

Proposition A.1 further shows that, for characteristic kernels, Q1(x0)=0Q_1(\boldsymbol{x}_0)=0 and Q2(x0)=0Q_2(\boldsymbol{x}_0)=0 each hold if and only if ρ1=ρ2\rho_1=\rho_2, while Q3(x0)=0Q_3(\boldsymbol{x}_0)=0 if and only if δ=0\delta=0. Thus, changes in the correlation between Y1Y_1 and Y2Y_2 contribute through their respective association components, whereas a marginal change in Y3Y_3 is attributed only to coordinate 3 through its marginal component. This illustrates how PIT standardization separates association changes from marginal changes in other coordinates.

This marginal standardization is the key to our coordinate-level decomposition: it separates changes in the target marginal distribution from changes in its cross-association with the remaining coordinates. The use of conditional PITs is related to the conditional-copula principle, while the signed-measure contrast, the coordinate-level no-contribution hypothesis, and the associated inferential procedure are formulated for the present problem. We then embed the marginal and cross-association contrasts in a direct-sum RKHS and define the criterion Qj(x0)Q_j(\boldsymbol{x}_0). Proposition 2.1 shows that this criterion is zero if and only if the localized no-contribution null holds. The cross-fitted one-step estimator and multiplier bootstrap account for estimation of the conditional distributions and yield the weighted-χ2\chi^2 limiting null distribution described in Theorems 4.1–4.2. These definitions and results are developed in Sections 2–4, with Gaussian illustrations in Appendix A.

AuthorsOpenReview ↗

Response to Reviewer HCnF (Part 2/2)

Q1. Marginal Standardization and the Cross-Association Measure

We appreciate this question. The proposed CjC_j is designed to distinguish association changes involving coordinate jj from changes in the marginal distributions of the other coordinates.

  • Marginal standardization. On the raw scale, F(ℓ)−Fj(ℓ)⊗F−j(ℓ)F^{(\ell)}-F_j^{(\ell)}\otimes F_{-j}^{(\ell)} is a natural measure of departure from independence, but it is affected by changes in the marginal distributions of the coordinates in Y−jY_{-j}. It can therefore change even when the dependence involving YjY_j is unchanged. We instead apply the population-specific conditional PITs Ur(ℓ)=Fr(ℓ)(Yr(ℓ))U_r^{(\ell)}=F_r^{(\ell)}(Y_r^{(\ell)}) and define Cj(ℓ)=G(ℓ)−U⊗G−j(ℓ)C_j^{(\ell)}=G^{(\ell)}-\mathcal{U}\otimes G_{-j}^{(\ell)}, where G(ℓ)G^{(\ell)} is the conditional distribution of the PIT vector and U\mathcal{U} is the common uniform marginal distribution. The reference measure U⊗G−j(ℓ)\mathcal{U}\otimes G_{-j}^{(\ell)} preserves the joint distribution of the remaining transformed coordinates while removing their association with the target. Thus Cj(ℓ)C_j^{(\ell)} compares cross-association after marginal standardization; it does not require the target to be independent of Y−jY_{-j}.

  • Reference measure. This use of marginal standardization follows the conditional-copula principle (e.g., Veraverbeke et al., 2011), while the signed-measure contrast and the coordinate-level null are formulated here. Appendix A makes the motivation explicit: in Proposition A.1, changing the correlation between coordinates 1 and 2 yields Q1,Q2>0Q_1,Q_2>0 while a shift in the independent third-coordinate marginal yields Q3>0Q_3>0 only. Proposition A.2 shows that the corresponding raw-scale association measure would falsely attribute that third-coordinate shift to coordinates 1 and 2. The PIT construction and coordinate-level contrast are in Section 2; the Gaussian illustrations and raw-scale counterexample are in Appendix A (Propositions A.1–A.2).

Q2. Why Separate Marginal and Cross-Association Contributions

We appreciate the proposed alternative. The full-minus-reduced distance and LTCC address different attribution questions: LTCC distinguishes changes in conditional marginal distributions from changes in marginal-standardized cross-association.

The proposed full-minus-reduced distance measures an incremental discrepancy: how much a chosen distance changes when coordinate jj is added to Y−jY_{-j}. LTCC targets a different estimand, namely whether coordinate jj has a different conditional marginal distribution or a different marginal-standardized cross-association with the remaining coordinates.

These targets can disagree. Let Z(ℓ)∼N(μℓ,1)Z^{(\ell)}\sim N(\mu_\ell,1), ε(ℓ)∼N(0,σ2)\varepsilon^{(\ell)}\sim N(0,\sigma^2) be independent, and set Y2(ℓ)=Z(ℓ)Y_2^{(\ell)}=Z^{(\ell)}, Y1(ℓ)=Z(ℓ)+ε(ℓ)Y_1^{(\ell)}=Z^{(\ell)}+\varepsilon^{(\ell)}, with μ1≠μ2\mu_1\ne\mu_2. Then Y1(ℓ)∣Y2(ℓ)=y∼N(y,σ2)Y_1^{(\ell)}\mid Y_2^{(\ell)}=y\sim N(y,\sigma^2) is common across populations, while F1(1)≠F1(2)F_1^{(1)}\ne F_1^{(2)} and the Gaussian copula is unchanged. LTCC therefore regards coordinate 1 as contributing through its marginal component. For the total variation distance, the common conditional kernel implies dTV(F(1),F(2))=dTV(F2(1),F2(2))d_{\rm TV}(F^{(1)},F^{(2)})=d_{\rm TV}(F_2^{(1)},F_2^{(2)}), so the full-minus-reduced increment is zero. The two constructions answer different attribution questions, and LTCC explicitly separates the marginal and cross-association components. Proposition 2.1 also shows why LTCC is a coherent decomposition: all coordinate criteria vanish exactly when the two conditional joint distributions agree. The decomposition and its exact characterization are given in Section 2 and Proposition 2.1; the Gaussian counterexample above illustrates the distinction between the two attribution targets.


We thank the reviewer again for these constructive questions, which helped us clarify the estimand and its interpretation.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Rebuttal Responses and Manuscript Revisions

We thank the reviewers and the Area Chair for their time and constructive feedback. During the rebuttal, we addressed the concerns summarized below and incorporated the new theoretical results, experiments, and computational analysis into the revised manuscript.

1. Research objective, practical importance, and relation to existing methods

We clarified LTCC’s research objective, the practical importance of dependence-only contributions, and the role of population-specific conditional PITs. We explained how LTCC differs from existing conditional distribution tests and predictive variable-importance methods, and clarified that the baseline comparisons illustrate different notions of coordinate relevance. We also strengthened the practical motivation in the Introduction. (HCnF: W, Q1–Q2; UCJ7: Q4–Q5; crvd: W1/Q1, Q6)

2. Connection between theory and finite Monte Carlo implementation

We added a Monte Carlo approximation-error analysis and sufficient conditions for preserving asymptotic calibration, together with a sensitivity study examining the effect of the conditional-sampling budget. These additions address the impact of finite Monte Carlo approximation on test validity and are included in Appendix F and Appendix G.1 (Table 5), respectively. (UCJ7: W1/Q1)

3. First-stage convergence conditions and performance with multiple covariates

We clarified the first-stage convergence conditions and provided a conditional-density estimation example showing how they can be satisfied under suitable regularity assumptions. We also clarified the dimensional assumptions of the theory and added a two-covariate experiment to evaluate empirical size and power when multiple covariates jointly affect the conditional distribution. This experiment has been incorporated into Appendix G.2 (Figure 8). (UCJ7: W2/Q2; crvd: W2/Q2, W3)

4. Multiple testing across coordinates and simultaneous inference across profiles

We discussed multiple testing across response coordinates and clarified its distinction from simultaneous inference across covariate profiles. For a fixed response coordinate over a fixed finite set of prespecified profiles, we added a joint max-bootstrap test with asymptotic level control and consistency in Appendix H. (UCJ7: W3, Q3; crvd: W4/Q4)

5. Local alternatives, asymptotic power, and kernel-smoothing conditions

We added a local-alternative analysis characterizing the asymptotic power and local detection scales of LTCC. We also clarified the relationship between kernel order, covariate dimension, and bandwidth conditions. The new theoretical results and proofs are included in Appendix I. (crvd: Q5)

6. Computational complexity

We added a computational-complexity analysis of LTCC, separating the first-stage estimation and sampling costs from the subsequent testing costs. The analysis identifies the dominant computational steps and provides an overall complexity bound, clarifying the dependence on sample size, the number of conditional Monte Carlo draws, and the number of bootstrap repetitions. This analysis is included in Appendix D.4. (crvd: Q3)

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.