acceptodds
Under review as a conference paper at ICLR 2027

Geometry-Limited Policy Contrasts for Distribution-Valued Outcomes

Abstract

Many interventions produce a distribution for each individual rather than a scalar response. Comparing only means or Wasserstein barycenters then discards population heterogeneity among individual distributions. We study population matching costs between the resulting policy-specific outer laws, allowing all matching mechanisms compatible with their marginals. Conditional Kantorovich duality gives pairwise sharp bounds. We construct a split-sample lower certificate by extending a finite transport dual to the full quantile cone. Under known assignment probabilities, global feasibility yields finite-sample validity uniformly over a policy class even when the outer-law learner is misspecified. Validity, however, need not imply information. With finitely many covariate cells and actions, one-dimensional outcomes, and cost , the minimax information loss over unrestricted quantile outcomes is of order for honest lower limits and point estimation. A cost-active high-probability quantization condition is sufficient for a polynomial upper rate. We also propagate uncertainty from finite within-individual samples. A semi-synthetic experiment built from real electricity-load quantile templates gives a positive simultaneous lower certificate when population heterogeneity changes despite equal scalar means and Wasserstein barycenters.

Primary area:
causal reasoning
Keywords:
nested Wasserstein distance, population matching, partial identification, uniform confidence bounds, random measures

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_uCyDOpenReview ↗
Rating
1
Confidence
1
Informativeness
1
Presentation
1
Soundness
3
Read review
Paper summary

The paper compares policies through the population laws of individual outcome distributions. It studies the sharp range of cross-person matching costs and constructs globally feasible dual certificates with simultaneous coverage. For exactly observed distributions and fixed finite covariate/action sets, it establishes logarithmic minimax rates for both endpoints when local and global cost orders match. It also studies structured outer laws, estimated propensities, and finite within-person samples.

Critical strengths
  1. The minimax lower bounds apply to every estimator of either scalar endpoint, extending previous results on slow empirical transport convergence. Matching upper bounds and the persistence of the barrier under uniformly bi-Lipschitz individual quantiles provide a useful account of what makes these comparisons difficult.

  2. The certificates handle global feasibility, directional discretization errors, and policy search. The paper credits the inherited feasible-dual principle and adds calibrated finite-cell AIPW and finite-bag corrections. Controlled comparisons separate learning, extension, and calibration errors, supporting the distinction between valid certificates and informative ones.

Critical weaknesses
  1. The writing of current draft need serious polish. I can barely comprehend the introduction due to lack of transition words or coherence. The figures appear to be directly generated by AI and are difficult to understand. The current draft is not appropriate for publishing.
  2. Statistical attainability does not establish affordable computation. Completed anchor sets can grow exponentially in inverse precision (E.5), while general convex transforms may require certified global optimization (E.6). K.3.6 demonstrates improved quadratic extensions, but leaves runtime scaling unclear. The main text should distinguish tractable costs from the general construction.
  3. The small calibration-to-learning-gap ratio emphasized in Figure 3a can coexist with wide certificates. At total n=65536, the equal-budget quadratic experiment retains 36.1% of the lower endpoint (Table 30); the unknown-propensity experiment has lower/upper ratios of 0.786/1.572 (Table 3). These are different designs, but both show why the main discussion should include endpoint-normalized precision alongside calibration overhead.
  4. The main text is dense and relies heavily on appendix references. It states the finite-design restrictions, but the connection between the geometric conditions and the rates is difficult to follow. A short sketch linking the hard codes and active dual sections to the lower and upper bounds would improve accessibility.
Questions for authors

For a representative nonlinear directional loss, how does certified-transform runtime vary with quantile resolution and the required objective precision?

6 comments
AuthorsOpenReview ↗

Response 1/4 to Reviewer uCyD

wk1

We revised the prose and transitions in Section 1. The activity example shows how equal means and Wasserstein barycenters can conceal differences across individuals. The revised text connects this example to the range of costs across compatible cross-person matchings and explains what governs coverage and endpoint accuracy. The contribution list now leads with the minimax result for endpoint accuracy.

We redrew Figures 1 and 2 and rewrote their captions.

Figure 1 (Section 1). From individual distributions to a matching interval. Three panels connect the activity example to the matching interval and show how global feasibility supports coverage and how outer-law geometry governs endpoint accuracy.

Figure 2 (Section 3). The inference workflow. The figure shows training, global certification, independent evaluation, and simultaneous calibration in sequence.

We also revised the caption of Figure 3, submitted PDF, p. 9 to explain the two panels on calibration overhead and normalized endpoint bounds.

wk2

Joint response to WK2 and Question 1. The explicit full-cone grid used to attain the accuracy in Eq. (16) grows exponentially in the inverse numerical tolerance. We benchmark an asymmetric squared loss, a nonlinear convex loss whose full-cone transform can be nonconvex. On known finite support, a direct support scan bypasses grid completion and yields informative lower certificates in both directions in 5.100–11.006 seconds on average across the three tested resolutions, including verification. For two special classes, the full-cone transform itself is tractable: a direct scan for directional linear losses and a linear program for the symmetric squared loss.

AuthorsOpenReview ↗

Response 2/4 to Reviewer uCyD

Training-size benchmark at fixed K=8K=8. With the asymmetric squared loss h(v)=2(v+)2+0.5((−v)+)2h(v)=2(v_+)^2+0.5((-v)_+)^2, we fit on nested sets of 512, 1,024, and 2,048 distinct curves per arm. The known frozen support used for scoring and the 65,536 independent evaluation draws per arm were fixed across training sizes. Relative LCB shortfall is 1−LCB/truth1-\mathrm{LCB}/\mathrm{truth}, with each direction's continuous endpoint as the denominator. Values closer to 0 indicate smaller error.

Distinct training curves per armRelative LCB shortfall 0→10\to1Relative LCB shortfall 1→01\to0Preprocessing and fitting (s)Evaluation and calibration (ms)Verification (ms)Total incl. verification (s), mean [min–max]Peak CPU process working set (MiB)
5120.17250.15833.3451.7165.73.412 [3.208–3.807]239.07
1,0240.14350.14145.2111.6279.35.291 [4.245–7.301]239.38
2,0480.12410.12455.5301.3453.35.585 [5.302–5.951]301.52

Stage times average three independently launched single-thread processes per training size and cover both policy directions. Verification checks the support-pair inequalities and compares both LCBs with their continuous truth brackets. Peak memory is the largest whole-process CPU working set across the three runs. The resolution benchmark below retimed all three settings in a separate batch with training fixed at 2,048 curves per arm.

Anonymized stage-by-stage runtime and peak CPU memory figure shows the timing breakdown for this training-size benchmark.

Resolution benchmark on known frozen support (Table R1). Each run uses the same frozen population, asymmetric squared loss h(v)=2(v+)2+0.5((−v)+)2h(v)=2(v_+)^2+0.5((-v)_+)^2, training and independent evaluation indices, and Bernstein calibration. We fit on all 2,048 distinct curves per arm and use the same 65,536 independent evaluation draws per arm. The target score takes the minimum over the known frozen source support. The experiment covers exactly two ordered policy pairs, 0→10\to1 and 1→01\to0. Their positive continuous population lower endpoints are approximately 0.00529622 and 0.02088609, respectively. Relative LCB shortfall is 1−LCB/truth1-\mathrm{LCB}/\mathrm{truth}, using each direction's continuous endpoint as the denominator. Values closer to 0 indicate smaller error.

KKBinned OT 0→10\to1Binned OT 1→01\to0Relative LCB shortfall 0→10\to1Relative LCB shortfall 1→01\to0Total through verification (s), mean [min–max]Peak CPU working set (MiB)
80.005066730.020026060.124080.124475.100 [4.894–5.365]348.21
160.005229720.020647840.094800.096567.345 [6.940–7.695]348.66
320.005277570.020821100.086240.0887411.006 [10.180–11.421]464.01

All three resolutions were timed anew in the same Conda environment using three independently launched single-thread processes per KK. Each time includes both directions and numerical verification. The memory column gives the largest CPU process working set across the three runs.

The binned OT values increase toward the continuous endpoints at the tested resolutions. From K=8K=8 to K=32K=32, the binned OT shortfall fell from 4.33% to 0.35% in the 0→10\to1 direction and from 4.12% to 0.31% in the 1→01\to0 direction. Increasing KK from 16 to 32 reduced relative LCB shortfall by 0.86 and 0.78 percentage points, respectively, and added about 3.66 s to mean total time. Mean preprocessing and fitting rose from 7.225 to 10.888 s and accounted for nearly all of the increase. These measurements apply to the frozen support, the two directions, and the three resolutions tested.

For Question 1, Appendix K.3.6 (Table 25) reports a separate full-cone certification experiment for the same loss. With a budget of 129 nodes per query, all 16 queries at K=2K=2 met objective bracket widths of 10−210^{-2} and 10−310^{-3} in 2.916 s and 3.085 s, respectively. At K=4K=4, the transform stage took 5.073 s and 8.229 s, with 14/16 and 3/16 queries meeting the respective targets. At K=8K=8, it took 13.312 s and 8.554 s, with 3/16 and 2/16 meeting them. For settings with unmet targets, elapsed times include budget-limited queries and cannot establish time to the requested precision. Table R1 instead measures the cost of varying KK for a finite support scan, which has no full-cone objective tolerance τ\tau.

AuthorsOpenReview ↗

Response 3/4 to Reviewer uCyD

Accuracy target and computational cost. Equation (16) bounds the expected worst-pair positive shortfall of the completed-anchor lower certificate. Under its assumptions, the bound separates four contributions:

\mathbb E\sup_{(\pi,\pi')\in\Pi^2} \bigl\\{\Theta^L_{c_h}(\pi,\pi')-L_\alpha^{\mathrm{comp}}(\pi,\pi')\bigr\\}_+ \\ \lesssim\\ & \underbrace{(\ell\kappa\underline e+1)^{-1/s}}_{\text{outer-law learning}} +\underbrace{K^{-1}+\eta+\varepsilon_{\mathrm{opt}}}_{\text{numerical error}}\\\\ &+\underbrace{q_{\mathrm{ev}}}_{\text{nuisance remainder}} +\underbrace{\frac{1}{\underline e} \sqrt{\frac{\log(N_\Pi\vee 2)+\log(2/\alpha)}{m}}}_{\text{evaluation and calibration}}. \end{aligned}$$ Here $\ell$ and $m$ are the training and evaluation sample sizes, $\kappa$ is the minimum cell mass, $\underline e$ is the overlap lower bound, $s$ is the quantization exponent, and $N_\Pi$ counts the distinct policy action vectors. For known-propensity IPW, $q_{\mathrm{ev}}=0$. The three numerical terms come from quantile discretization ($K^{-1}$), the source-anchor covering radius ($\eta$), and the certified residual of the finite training OT solve ($\varepsilon_{\mathrm{opt}}$). Training anchors already give a globally feasible lower dual. Completion improves its tightness. **The explicit completion grid grows exponentially in $1/t$.** Here $t$ is the numerical error budget. Taking $t$ at the scale of the statistical error in Eq. (16) preserves the statistical rate. For fixed outcome range $D>0$ and loss Lipschitz constant $L_h>0$, the choices $K=\lceil16L_hD/t\rceil$, $\eta=t/(8L_h)$, and $\varepsilon_{\mathrm{opt}}\le t/2$ ensure $4L_hD/K+2L_h\eta+\varepsilon_{\mathrm{opt}}\le t$. Both $K$ and the number of grid intervals per coordinate, $\lceil D/\eta\rceil$, then grow as $1/t$. The number of nondecreasing grid vectors is $$J_{K,\eta}=\binom{K+\lceil D/\eta\rceil}{K} =\exp\\{\Theta(1/t)\\}\qquad (t\downarrow0).$$ Appendix E.5 constructs and scores every one of these anchors. Scanning $J_{\mathrm{fit}}$ fitted target sections per anchor costs $O(J_{K,\eta}J_{\mathrm{fit}}K)$ in total. The anchors carry zero transport mass, so the exponential growth is in the completion step. **Cost of the transform by loss and support.** For a query $r$, the full-cone transform minimizes $c_K(x,r)-A(x)$ over the monotone quantile cone, where $A(x)=\min_j\\{c_K(x,y_j)-\beta_j\\}$ is the fitted source potential. - **General convex loss on the full cone.** For this separate transform, Appendix E.6 uses branch-and-bound on a potentially nonconvex objective. The lower endpoint of a bracket of width $\tau$ gives a feasible score with objective error at most $\tau$. The number of boxes explored depends on $K$, $\tau$, and the instance. - **Directional linear loss.** The directed triangle inequality makes the transform equal to $-A(r)$, computed in $O(J_{\mathrm{fit}}K)$ per query. - **Symmetric squared loss.** The common quadratic term cancels, giving an LP with $K+1$ variables and $J_{\mathrm{fit}}+O(K)$ constraints per query. - **Known finite support, as in the benchmark.** For encoded support sizes $s_0,s_1$, building a directional cost matrix takes $O(Ks_0s_1)$ time and $O(s_0s_1)$ memory. After the separate empirical OT fit, compute $g(y)=\min_{x\in S_0}\\{c_K(x,y)-f(x)\\}$ for all target atoms in $O(s_0s_1)$ time with the cached matrix. The minimum enforces feasibility for all support pairs. The implementation applies a $2\times10^{-10}$ downward guard to each score. The full-support oracle fits every frozen curve at its population weight, so its training and population laws coincide. In both directions its population dual mean is within about $2.0\times10^{-10}$ of the binned OT value. Dual feasibility and conservative binning (Theorem E.1) bound the population score by the continuous endpoint. Binning can leave a gap to that endpoint, and the final LCB subtracts an empirical Bernstein radius from the independently evaluated score. Conditional on fitting, empirical Bernstein calibration on the independent draws gives at least 99% simultaneous coverage for the two directions at each training size by allocating a failure probability of 0.0025 to each of the four arm means. We propose stating the loss and support conditions for these computational routes after Eq. (16) and adding Table R1 to the experimental discussion, with a reference to the separate full-cone results in Appendix K.3.6.
Reviewer_9RzQOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper studies how to compare treatment policies based on each individual's outcome distributions. Specifically, it studies population matching costs between policy-specific outer laws, allowing all matching mechanisms compatible with the conditional marginals. The conditional Kantorovich duality characterizes the sharp lower and upper endpoints, and the paper provides a logarithmic minimax convergence rates, under the outer law.

Critical strengths
  1. The paper is well-motivated. Means and Wasserstein barycenters can overlook differences in population heterogeneity, and the proposed outer-law comparison captures these information.
  2. This is also a theoretically strong paper that establishes the minimax rates that characterizes the statistical difficulty of the problem.
Critical weaknesses
  1. The paper compares distributions across potentially different individuals under two policies, using arbitrary couplings of their outer laws. However, it remains unclear why the minimum or maximum across all possible hypothetical matchings should be the primary quantity used for policy decisions.
  2. The provided theoretical results rely on the critical assumptions that the covariates are finite.
  3. The proposed algorithm seems to be not computationally tractable. The authors should shed light on this issue.
  4. The principal theoretical results also treat each individual's outcome distribution as exactly observed, which is usually not the case in reality.
Questions for authors

Please refer to weakness.

5 comments
AuthorsOpenReview ↗

Response 1/4 to Reviewer 9RzQ

WK1. Why matching bounds matter for policy decisions

Jin, Ren, Yang, and Wang (2025, Section 3.2) provide a precedent for using a lower bound as a policy-selection criterion. Their pessimistic policy learning rule maximizes a lower confidence bound on policy value, so a policy is selected for the value the data can support. The same decision principle motivates using the lower endpoint when the objective is to maximize guaranteed benefit. Jin et al.'s bound addresses sampling uncertainty. Our population endpoints address uncertainty about how individuals are matched across policy populations.

Under our assumptions, the data identify each policy's conditional outer law, the distribution of individual outcome distributions given covariates. The cross-person coupling of these laws can remain unknown even with infinite data. If an application has an actual matching mechanism, the expected cost under that mechanism is the ground-truth contrast for the policy pair. Theorem 2.1 shows that this contrast lies between the population endpoints and that every value in the interval is attained by an admissible matching. Thus the coupling-dependent contrast is partially identified, and the endpoints are the strongest bounds available from the outer laws alone. The finite-sample certificates in Section 3 account for sampling uncertainty about these endpoints.

Which endpoint to use depends on the decision. With a directional loss measuring excess harm relative to a fixed baseline, the lower endpoint gives the least harm possible under any admissible matching, and the upper endpoint gives the most. A decision maker who wants to minimize worst-case harm would choose the policy with the smallest upper endpoint. If the lower endpoint exceeds a safety threshold, every admissible matching violates it. If the upper endpoint is at or below the threshold, every admissible matching satisfies it. For example, with harm intervals of [0.1,0.6][0.1,0.6] and [0.2,0.3][0.2,0.3], a best-case rule selects the first policy, while a worst-case rule selects the second. An interior value is attainable under a compatible matching and can inform a scenario-specific decision when that matching is substantively justified. The selected policy may then depend on the matching assumed.

The revised Introduction now makes explicit that a symmetric cost measures population discrepancy and that a directional cost can encode the harm and tolerance relevant to a particular decision.

WK2. Finite covariate support

The minimax rate extends to continuous covariates. The submitted sharpness and finite-sample IPW coverage results already allow continuous covariates.

New corollary. Continuous-design minimax rate. Use the known-propensity setting of Theorem 4.1, with matching cost orders βh=γh=β\beta_h=\gamma_h=\beta and a fixed finite policy class containing two distinct constant policies. Replace the finite covariate support by X∈[0,1]dX\in[0,1]^d for fixed dd, with density bounded above and away from zero. Assume the conditional outer laws satisfy, uniformly over actions,

where W1QW_1^{\mathcal Q} uses the L1L^1 distance between quantile functions. Then both sharp endpoints retain the minimax rate

for point estimation and simultaneous honest one-sided inference at any fixed 0<α<1/30<\alpha<1/3.

Proof sketch. Fit duals in shrinking covariate bins and evaluate them by independent IPW at the original covariates for the upper rate. Conditional outer laws constant in xx give the matching lower rate.

AuthorsOpenReview ↗

Response 2/4 to Reviewer 9RzQ

WK3. Computational tractability

On the known frozen support, a direct support scan gave informative lower confidence bounds in both policy directions at all three tested resolutions, with mean runtimes of 5.100–11.006 seconds including verification. The full-cone construction has a different computational cost. The explicit grid in Appendix E.5 grows exponentially in inverse numerical tolerance, and Appendix E.6 gives simpler transforms for particular losses.

Training-size benchmark at fixed K=8K=8. With the asymmetric squared loss h(v)=2(v+)2+0.5((−v)+)2h(v)=2(v_+)^2+0.5((-v)_+)^2, we fit on nested sets of 512, 1,024, and 2,048 distinct curves per arm. The known frozen support used for scoring and the 65,536 independent evaluation draws per arm were fixed across training sizes. LCB/truth uses each direction's continuous endpoint as the denominator.

Distinct training curves per armLCB/truth 0→10\to1LCB/truth 1→01\to0Preprocessing and fitting (s)Evaluation and calibration (ms)Verification (ms)Total incl. verification (s), mean [min–max]Peak CPU process working set (MiB)
5120.82750.84173.3451.7165.73.412 [3.208–3.807]239.07
1,0240.85650.85865.2111.6279.35.291 [4.245–7.301]239.38
2,0480.87590.87555.5301.3453.35.585 [5.302–5.951]301.52

Stage times average three independently launched single-thread processes per training size and cover both policy directions. Verification checks the support-pair inequalities and compares both LCBs with their continuous truth brackets. Peak memory is the largest whole-process CPU working set across the three runs. The resolution benchmark below retimed all three settings in a separate batch with training fixed at 2,048 curves per arm.

Anonymized stage-by-stage runtime and peak CPU memory figure shows the timing breakdown for this training-size benchmark.

Resolution benchmark at fixed training size. We varied K∈8,16,32K\in\\{8,16,32\\} while holding fixed the uniform frozen population, the asymmetric squared loss h(v)=2(v+)2+0.5((−v)+)2h(v)=2(v_+)^2+0.5((-v)_+)^2, the training and independent evaluation indices, and the empirical Bernstein calibration. All three settings fit on all 2,048 distinct frozen curves per arm and use 65,536 independent evaluation draws per arm. Each fitted target score is a direct minimum over the known frozen source support. Because the training and population laws coincide, the population dual means match the binned OT references to within about 2.0×10−102.0\times10^{-10}. The final bounds still reflect conservative binning and independent evaluation with calibration. LCB/truth divides the lower confidence bound by the original continuous population truth for its direction.

KKLCB/truth 0→10\to1LCB/truth 1→01\to0Total time, s (mean [min–max])Peak CPU process working set (MiB)
80.875920.875535.100 [4.894–5.365]348.21
160.905200.903447.345 [6.940–7.695]348.66
320.913760.9112611.006 [10.180–11.421]464.01

Total time sums the measured stages through verification for the two ordered policy pairs, 0→10\to1 and 1→01\to0. Each KK was timed in three freshly launched single-thread processes in the same Conda environment. Peak memory is the largest whole-process CPU working set across the three runs.

Increasing KK from 16 to 32 improves LCB/truth by 0.86 and 0.78 percentage points in the two respective directions. Mean total time increases by about 3.66 s, driven by preprocessing and fitting (7.225 to 10.888 s), chiefly by construction of the directional cost matrices. Independent evaluation and calibration average 1.47–1.70 ms across settings. Verification averages 116–124 ms and reuses the cached matrices to check support-pair feasibility and compare both LCBs with their precomputed continuous truth brackets. Each LCB is below the lower endpoint of its truth bracket. The comparisons are numerical checks. Support feasibility and independent empirical Bernstein calibration give at least 99% simultaneous coverage for the two directions at each fixed KK, allocating error 0.00250.0025 to each of the four score means. The measured trade-off applies to this frozen support, these two directions, and the three tested resolutions.

AuthorsOpenReview ↗

Response 3/4 to Reviewer 9RzQ

Accuracy target and computational cost. Equation (16) bounds the expected worst-pair positive shortfall of the completed-anchor lower certificate. Under its assumptions, the bound separates four contributions:

\mathbb E\sup_{(\pi,\pi')\in\Pi^2} \bigl\\{\Theta^L_{c_h}(\pi,\pi')-L_\alpha^{\mathrm{comp}}(\pi,\pi')\bigr\\}_+ \\ \lesssim\\ & \underbrace{(\ell\kappa\underline e+1)^{-1/s}}_{\text{outer-law learning}} +\underbrace{K^{-1}+\eta+\varepsilon_{\mathrm{opt}}}_{\text{numerical error}}\\\\ &+\underbrace{q_{\mathrm{ev}}}_{\text{nuisance remainder}} +\underbrace{\frac{1}{\underline e} \sqrt{\frac{\log(N_\Pi\vee 2)+\log(2/\alpha)}{m}}}_{\text{evaluation and calibration}}. \end{aligned}$$ Here $\ell$ and $m$ are the training and evaluation sample sizes, $\kappa$ is the minimum cell mass, $\underline e$ is the overlap lower bound, $s$ is the quantization exponent, and $N_\Pi$ counts the distinct policy action vectors. For known-propensity IPW, $q_{\mathrm{ev}}=0$. The three numerical terms come from quantile discretization ($K^{-1}$), the source-anchor covering radius ($\eta$), and the certified residual of the finite training OT solve ($\varepsilon_{\mathrm{opt}}$). Training anchors already give a globally feasible lower dual. Completion improves its tightness. **The explicit completion grid grows exponentially in $1/t$.** Here $t$ is the numerical error budget. Taking $t$ at the scale of the statistical error in Eq. (16) preserves the statistical rate. For fixed outcome range $D>0$ and loss Lipschitz constant $L_h>0$, the choices $K=\lceil16L_hD/t\rceil$, $\eta=t/(8L_h)$, and $\varepsilon_{\mathrm{opt}}\le t/2$ ensure $4L_hD/K+2L_h\eta+\varepsilon_{\mathrm{opt}}\le t$. Both $K$ and the number of grid intervals per coordinate, $\lceil D/\eta\rceil$, then grow as $1/t$. The number of nondecreasing grid vectors is $$J_{K,\eta}=\binom{K+\lceil D/\eta\rceil}{K} =\exp\\{\Theta(1/t)\\}\qquad (t\downarrow0).$$ Appendix E.5 constructs and scores every one of these anchors. Scanning $J_{\mathrm{fit}}$ fitted target sections per anchor costs $O(J_{K,\eta}J_{\mathrm{fit}}K)$ in total. The anchors carry zero transport mass, so the exponential growth is in the completion step. **Cost of the transform by loss and support.** For a query $r$, the full-cone transform minimizes $c_K(x,r)-A(x)$ over the monotone quantile cone. Here $c_K(x,r)=K^{-1}\sum_{k=1}^K h(r_k-x_k)$ is the binned cost and $A(x)=\min_j\\{c_K(x,y_j)-\beta_j\\}$ is the fitted source potential. - **General convex loss on the full cone.** For this separate transform, Appendix E.6 uses branch-and-bound on a potentially nonconvex objective. The lower endpoint of a bracket of width $\tau$ gives a feasible score with objective error at most $\tau$. The number of boxes explored depends on $K$, $\tau$, and the instance. - **Directional linear loss.** The directed triangle inequality makes the transform equal to $-A(r)$, computed in $O(J_{\mathrm{fit}}K)$ per query. - **Symmetric squared loss.** The common quadratic term cancels, giving an LP with $K+1$ variables and $J_{\mathrm{fit}}+O(K)$ constraints per query. - **Known finite support, as in the benchmark.** For encoded support sizes $s_0,s_1$, building a directional cost matrix takes $O(Ks_0s_1)$ time and $O(s_0s_1)$ memory. After the separate empirical OT fit, compute $g(y)=\min_{x\in S_0}\\{c_K(x,y)-f(x)\\}$ for all target atoms in $O(s_0s_1)$ time with the cached matrix. The minimum enforces feasibility for all binned support pairs, and conservative binning transfers the inequality to the original curves. The implementation applies a $2\times10^{-10}$ downward guard to each score. We propose adding the resolution comparison and its scope to the computational discussion in Appendix K and retaining the distinction between statistical precision and runtime next to Eq. (16).
Reviewer_jHSbOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
1
Soundness
3
Read review
Paper summary

The paper compares treatment policies when each person's outcome is a distribution rather than a single number. It studies the range of costs obtained by matching individuals across the two policy populations, under the stated causal assumptions. The main theoretical results explain why estimating this range can remain difficult even when individual distributions are smooth. The paper also constructs confidence bounds that remain valid with imperfect initial estimates, under the stated feasibility and evaluation conditions, and considers extensions for unknown assignment probabilities, numerical approximation, and repeated measurements.

Critical strengths

An interesting problem beyond average outcomes. Sections 1-2 explain why policies can produce populations with the same average outcome and Wasserstein barycenter while differing in the variation between individuals' distributions. Studying the range of matching costs offers a meaningful way to retain this information. Valuable results on the limits of estimation. Section 4 shows that estimating either endpoint of the matching range can remain very slow even when individual distributions are smooth. The lower and upper accuracy bounds help explain when this difficulty is intrinsic to the problem and when additional population structure can improve estimation (Theorems 4.1-4.2 and Corollary 4.3). A useful distinction between validity and precision. Section 3 explains why confidence bounds can remain valid despite inaccurate initial estimates, under the stated conditions, while Section 4 studies how close they can become to the target endpoints. The underlying duality-based inference has precedents, as the paper acknowledges; connecting it to the geometric analysis and explicit error accounting is valuable.

Critical weaknesses

The presentation makes the central contribution too difficult to assess. Although the scientific main text is nine pages, the complete manuscript runs to 119 pages. The main text compresses many results, extensions, and technical conditions, with substantial supporting material needed to understand their scope. Even with the overview figures and proof guide, it remains difficult to distinguish the essential argument from the additional developments and to track the assumptions behind each claim. This is the main reason for my recommendation: the paper needs a substantially more focused presentation, with the central results and their implications made clear in the main text. The practical computing requirements need a clearer account. Equation (16) explains how numerical errors affect the guarantee, but it is less clear what resources are required to make those errors small enough for useful bounds. The quadratic-cost experiments provide helpful evidence for a specific implementation. Existing timing and memory measurements, with preprocessing reported separately, would clarify its practical reach (Appendices E and K.3.6). The empirical support has a limited scope. Section 5 provides useful controlled tests, including electricity-load templates, but leaves open how informative the bounds would be across more varied populations. The main analysis assumes fixed finite designs and fully observed individual distributions, as acknowledged in Section 6. These qualifications limit the demonstrated practical reach of the method; the strongest case for the paper remains its theoretical insight.

Questions for authors

Core contribution and assumptions. Could you give a concise account of which results are essential to the main contribution and the assumptions each requires? In particular, clarifying what the geometric analysis adds to the existing duality-based inference, and how the main experiments illustrate those additions, would help assess the paper's contribution. Computational cost. What do the existing implementations establish about the resources needed to obtain informative bounds under equation (16)? Could you summarize running time and peak memory for the reported experiments, separating preprocessing from subsequent evaluation and including the work required to check the guarantees across policy pairs (Appendix K, Tables 28-30)? Reproducibility. The manuscript mentions an archive in Appendix K.5. Could you indicate where it can be accessed, potentially through an anonymized repository? This would help verify the data preprocessing and rerun the experiments in Section 5 and Figure 3 with the original implementation.

8 comments
AuthorsOpenReview ↗

Response 1/6 to Reviewer jHSb

Presentation revisions

The Introduction now leads with the geometric result for both sharp matching endpoints, including the logarithmic minimax rate that persists for smooth individual distributions when the two cost orders match. It explains how feasible-dual inference supports coverage. In Section 4, Table 1 states the assumptions for sharpness, coverage and minimax accuracy, and the text defines the risks and sketches the proof of Theorem 4.1. Section 5 connects the main experiments to calibration, learning error and between-person heterogeneity. The supplement's reading guide points to Appendix F for the core geometric proofs.

Core contribution and assumptions

The policy comparison and its assumptions

How much discrepancy between two policy populations is unavoidable when the pairing of people is unspecified? With wearable activity data, equal and opposite shifts for two halves of a homogeneous population preserve the mean and Wasserstein barycenter. Every pairing with the original population still incurs a positive Wasserstein cost. Theorem 2.1 gives the sharp range over within-covariate matchings. It uses bounded continuous costs on a compact outcome space and identifies the marginals under consistency, no interference, conditional exchangeability and overlap. Cell counts and assignment records can flag weak overlap and motivate restricting the target to supported actions. Unmeasured confounding would require sensitivity assumptions on the marginals.

What dual feasibility guarantees

Global dual feasibility gives conservative population bounds under outer-law misspecification. A poor fit can leave a large gap even with exact evaluation. We adapt the feasible-dual approach of Ji, Lei, and Spector (2023) by extending finite duals to the full quantile cone with separate discretization corrections for each endpoint. Given overlap, bounded feasible potentials and independent evaluation, simultaneous coverage uses IPW with known propensities or clipped AIPW with a certified finite-cell nuisance remainder (Theorems 3.1–3.2 and Corollary 3.3). A certified bound on constraint violations can be absorbed by widening the interval.

Figure 3(b) shows a positive sharp W22W_2^2 lower endpoint for electricity-load template populations with equal barycenters. The intervals from both fits covered the two endpoints in all 1,000 repetitions per sample size. At n=20,000n=20{,}000, the mean lower bound retained 96.3% of the sharp value under the correct fit and 83.1% under misspecification.

Figure 3(b), submitted PDF, p. 9. Bounds relative to the sharp endpoints under the correct and misspecified fits.

What the geometric analysis adds

Catalano and Lavenant (2024, Section 4) show slow convergence of empirical outer laws in Wasserstein-over-Wasserstein distance. A scalar functional can be easier to estimate than an entire law. Theorems 4.1–4.2 give lower bounds for every estimator and honest band for either matching endpoint, plus attainable upper rates. These rates assume a fixed finite design with known covariate support, cell masses and propensities bounded below, at least two distinct policy action vectors, and exact individual distributions on a compact real interval. IPW coverage allows general covariates.

For a bounded convex displacement loss hh, local contact of order βh\beta_h gives a lower bound on departure from a supporting line. We use it to separate exponentially large quantile codes and reduce endpoint estimation to two-sample closeness testing. A separate code makes mismatches cheaper for the upper endpoint. This yields a lower bound of order (log⁡n)−βh(\log n)^{-\beta_h} for both endpoints. A global rectangular bound of order γh\gamma_h on mixed differences of hh limits the active cost sections in a dual approximation. The resulting entropy bound gives an attainable rate of (log⁡n)−γh(\log n)^{-\gamma_h}. Each condition is checkable from the chosen loss. Uniformly approximating it within ϵ\epsilon moves each endpoint by at most ϵ\epsilon (Appendix G). Revised Table 1 maps assumptions to results.

AuthorsOpenReview ↗

Response 2/6 to Reviewer jHSb

Minimax rates and evidence from Figure 3

When βh=γh\beta_h=\gamma_h, point estimation and simultaneous honest inference share the minimax rate (log⁡n)−βh(\log n)^{-\beta_h} at both endpoints because calibration terms decay faster. For WppW_p^p the exponent is p∧2p\wedge2, even with uniformly bi-Lipschitz quantiles (Corollary 4.3). The loss can distinguish many smooth curves, and the lower bound already holds for a fixed nontrivial policy pair with known propensities. On growing smooth W1W_1 codes, the IPW correction fell from 37.1% to 8.4% of the mean maximal learning gap (Figure 3(a)). At n=12,288n=12{,}288, AIPW's total correction was 8.0% of its mean maximal learning gap, yet its mean lower bound for a fixed contrast retained only 16.7% of the sharp value (Table 14).

Figure 3(a), submitted PDF, p. 9. Calibration cost relative to the mean maximal learning gap.

When population structure improves precision

Appendix I gives a polynomial upper rate under the finite design when polynomial-size codebooks approximate the conditional populations in both cost-induced pseudometrics, with sufficiently fast decay of uncovered mass. This gives policy learning a criterion for choosing representations that preserve matching costs, including directional losses for threshold exceedances. Held-out quantization errors can assess the approximation at observed scales. Appendix J gives coverage corrections for conditionally i.i.d. finite bags with known propensities and suitable Lipschitz potentials. Fixed bag sizes can leave the latent endpoints unidentified, raising the question of how to allocate measurements across and within people.

Informativeness across more varied populations

What the theory says about informativeness

The worst-case precision of a valid certificate improves only logarithmically, even for smooth individual distributions. The population experiments below measure how close the certificates get to the endpoints in five specific designs.

Joint coverage and endpoint accuracy under a fixed finite design.

Setup. Observe individual outcome distributions on [0,1][0,1] exactly. Assume consistency, no interference, conditional exchangeability, and overlap ea(x)≥e‾>0e_a(x)\ge\underline e > 0. The fixed, known covariate support has KXK_X cells with masses at least κ>0\kappa > 0, there are dd actions, and the policies induce at least two distinct action vectors. Use ch(q,r)=∫01h(r(v)−q(v)),dvc_h(q,r)=\int_0^1h(r(v)-q(v))\\,dv for bounded convex h:[−1,1]→[0,hˉ]h:[-1,1]\to[0,\bar h] with h(0)=0h(0)=0. For u∈Π2u\in\Pi^2, let TPL(u)T_P^L(u) and TPU(u)T_P^U(u) be the minimum and maximum expected matching cost over compatible within-covariate matchings.

Coverage. With known propensities, fit bounded, globally feasible duals for chc_h and its complement hˉ−ch\bar h-c_h. Use independent IPW evaluation with pointwise measurable score classes and calibrate each lower certificate at level 1−α/21-\alpha/2. Subtract the complement certificate from hˉ\bar h to obtain UU. The resulting bands satisfy, even under outer-law misspecification,

\text{ for all }u\in\Pi^2\\}\ge1-\alpha.

Accuracy. Measure risk by the worst-case expected maximum over policy pairs of ∣T^e−TPe∣|\widehat T^e-T_P^e| for estimation or (TPL−L)+(T_P^L-L)_+ and (U−TPU)+(U-T_P^U)_+ for honest inference. If h(z0+v)−h(z0)−gv≥c∣v∣βhh(z_0+v)-h(z_0)-gv\ge c|v|^{\beta_h} locally for some interior z0z_0, subgradient gg, and c>0c > 0, all four risks are at least of order log⁡(n+1)−βh\\{\log(n+1)\\}^{-\beta_h} at fixed 0<α<1/30 < \alpha < 1/3. For each honest band, some law also gives a gap of this order with probability at least 1/31/3.

For the upper bound, assume hh is (γh∧1)(\gamma_h\wedge1)-Hölder and its mixed difference ∣h(z+v+w)−h(z+v)−h(z+w)+h(z)∣|h(z+v+w)-h(z+v)-h(z+w)+h(z)| is bounded by a constant times min⁡(∣v∣γh,∣w∣γh)\min(|v|^{\gamma_h},|w|^{\gamma_h}) for 0<γh≤10 < \gamma_h\le1, or min⁡(∣v∣∣w∣γh−1,∣w∣∣v∣γh−1)\min(|v||w|^{\gamma_h-1},|w||v|^{\gamma_h-1}) for 1<γh≤21 < \gamma_h\le2, whenever all four arguments are in [−1,1][-1,1]. Split-sample empirical duals then attain a point-estimation risk bounded by

\\{\log(\ell\kappa\underline e+1)\\}^{-\gamma_h} +q_{\rm ev} +\frac{\bar h}{\underline e}\sqrt{\frac{K_X\log d}{m}} \right).

Here ℓ\ell and mm are the training and evaluation sizes, and qev=0q_{\rm ev}=0 for known-propensity IPW or O(log⁡(n)/t)O(\log(n)/t) for finite-cell AIPW with an independent nuisance fold of size tt. Honest bands add log⁡(1/α)\log(1/\alpha) inside the square-root numerator. With fixed-proportion splits and βh=γh\beta_h=\gamma_h, the four risks all have minimax rate (log⁡n)−βh(\log n)^{-\beta_h}. The same rate holds if individual quantiles are restricted to a fixed Lipschitz class. For WppW_p^p, p≥1p\ge1, the exponent is p∧2p\wedge2, giving (log⁡n)−2(\log n)^{-2} for the squared W2W_2 cost.

AuthorsOpenReview ↗

Response 3/6 to Reviewer jHSb

The submitted proofs are in Sections 3–4 and Appendix F.

Two-group, four-group, and rare-type populations

We tested two- and four-group mixtures and three populations with rare types. The ratios below measure precision for the main pair, while joint coverage counts a draw only when both endpoints are covered for all four prespecified pairs.

Population4,096: main-pair LCB / true lower4,096: main-pair UCB / true upper4,096: four-pair joint coverage65,536: main-pair LCB / true lower65,536: main-pair UCB / true upper65,536: four-pair joint coverage
Two group0.74901.1036200/2000.94801.0199200/200
Four group0.81901.1090200/2000.96231.0203200/200
Rare 1%0.85721.1166200/2000.97851.0181200/200
Rare 5%0.81421.1360200/2000.96881.0248200/200
Rare 10%0.78481.1395200/2000.96201.0261200/200

Table note. The main pair (0000,1111)(0000,1111) has the same true lower endpoint, 0.0002256141264508190.000225614126450819, in all constructions. We use squared W2W_2 on a known template segment, with ⌊0.2n⌋\lfloor0.2n\rfloor training observations and the remaining approximately 80% for independent evaluation. Each population and budget has five constructions and 40 draws per construction, for 200 draws in total. Ratios are averaged within each construction and then equally across constructions. The nominal 95% joint intervals protect (0000,1111)(0000,1111), (0000,0011)(0000,0011), (0000,1100)(0000,1100), and (0011,1100)(0011,1100).

At n=65,536n=65{,}536, the mean main-pair lower bounds retained 94.8% to 97.9% of the true lower endpoint, and the mean upper bounds were 1.8% to 2.6% above the true upper endpoint. These averages quantify precision in the five controlled designs at a common signal and with known segment support.

Computational cost

Symmetric squared loss (submitted Appendix K, Tables 28–30)

We replayed the original configuration with d=K=8d=K=8, four cells, 16 policies, and 256 ordered policy pairs. The replay used one fixed statistical draw from the original seed at each training size and three single-thread timing runs of that draw on an Intel Core i7-1260P. The table gives mean times, with the minimum and maximum in brackets for the total. Peak memory is the largest whole-process CPU working set across the three runs.

StageTraining ℓ=4,096\ell=4{,}096Training ℓ=32,768\ell=32{,}768
Input loading and preprocessing (s)0.30800.3325
Empirical OT fitting (s)0.15573.2970
Setup of eight directional models (s)0.06690.7193
Certified transforms and cached scores (s)21.349213.3371
Evaluation, calibration and truth checks for all 256 pairs at both budgets (ms)5.056.80
Frozen-support feasibility and numerical checks (s)0.44640.4948
Continuous pipeline total (s)22.331 [19.048–24.726]18.187 [16.523–19.835]
Peak CPU process working set (MiB)286.00327.80

Evaluation reuses the fitted score tables for m=ℓm=\ell and m=64ℓm=64\ell. It computes the simultaneous Bernstein lower bounds and compares all 256 bounds with their frozen-population truths at each budget. The separate verification stage checks the directional score inequalities on the frozen support, the transform brackets, and the empirical OT objective. At ℓ=32,768\ell=32{,}768, increasing mm from ℓ\ell to 64ℓ64\ell raised the fixed-pair LCB from 32.21% to 81.84% of its truth on this draw. Both budgets covered all 256 pairs on that draw. The calibration argument supplies the statistical guarantee under the stated assumptions.

The pipeline total sums the listed stages and uses a precomputed truth reference. Building that reference separately took 5.335 s and peaked at 262.58 MiB. The whole-worker peak also includes the optional completed-anchor comparator, archive checks and the Python runtime.

Asymmetric squared loss on known frozen support

Training-size benchmark at fixed K=8K=8. With the asymmetric squared loss h(v)=2(v+)2+0.5((−v)+)2h(v)=2(v_+)^2+0.5((-v)_+)^2, we fit on nested sets of 512, 1,024, and 2,048 distinct curves per arm. The known frozen support used for scoring and the 65,536 independent evaluation draws per arm were fixed across training sizes. LCB/truth uses each direction's continuous endpoint as the denominator.

Distinct training curves per armLCB/truth 0→10\to1LCB/truth 1→01\to0Preprocessing and fitting (s)Evaluation and calibration (ms)Verification (ms)Total incl. verification (s), mean [min–max]Peak CPU process working set (MiB)
5120.82750.84173.3451.7165.73.412 [3.208–3.807]239.07
1,0240.85650.85865.2111.6279.35.291 [4.245–7.301]239.38
2,0480.87590.87555.5301.3453.35.585 [5.302–5.951]301.52

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.