acceptodds
Under review as a conference paper at ICLR 2027

Preserve the Target, Relax the Sharpness: Certified Compression for Scalable Instrumental Variable Partial Identification

Abstract

Sharp partial identification protects against false certainty, but exact causal representations can become large as outcome support grows. Naive coarsening reduces computation by changing the observed representation and can silently change the estimand or exclude data-compatible causal effects. We introduce a different computational primitive for finite instrumental-variable (IV) models: certified target-preserving stochastic representation reduction. A column-stochastic outcome map is lifted independently to every potential-outcome coordinate while treatment-response types are left unchanged. If the map preserves outcome values in expectation, observation commutes with the lift, every potential-outcome mean and affine mean contrast is preserved model-by-model, treatment-response restrictions are retained, and the original identified set is contained in the compressed set for arbitrary finite instrument and treatment cardinalities. Within the shared product-kernel class we characterize necessity and local uniqueness; with endpoint retention, nontrivial deterministic support reduction cannot satisfy the same exact mean-preservation certificate, so stochasticity is structural rather than cosmetic. We prove conditional sharpness bounds and a negative theorem excluding any support-uniform mesh-only rate, then turn nested support refinement into an anytime certified algorithm whose exact surrogate sets tighten monotonically to the original finite-support problem. A running-intersection rule preserves this monotonicity when the downstream causal optimizer is itself only outer-certified. Empirically, containment holds in all 13,660 generalized validation cells and all 7,692 nested refinement states. Adaptive allocation gives modest gains on nonadversarial cases and larger gains on an adversarial family, where observable-displacement-guided refinement reduces median width inflation at retained support size from 7.04% to 5.90%. Fresh-process one-thread timing shows representative multivalued-IV speedups of –, while an original outcome support of size compressed to gives a median speedup against the fastest certified exact backend. A nonasymptotic finite-sample extension and a public Job Training Partnership Act (JTPA) benchmark retain the same one-sided validity. We do not evade the exponential barrier for exact sharp analytical IV representations; we change the objective so that computation is traded against sharpness, not against the scientific target or validity.

Primary area:
causal reasoning
Keywords:
causal inference, partial identification, instrumental variables, causal reasoning, uncertainty quantification, certified compression, causal abstraction

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_2ne5OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a certified compression framework for finite instrumental-variable (IV) partial identification. The main idea is to reduce the outcome support using a stochastic mapping that preserves the causal target while allowing the identified set to become less sharp. The authors establish theoretical guarantees for target preservation and outer validity, introduce an anytime refinement procedure, and demonstrate substantial computational speedups on synthetic experiments and the JTPA dataset.

Critical strengths

The theoretical framework is generally well developed. The main results are supported by explicit constructions and proofs, including a useful counterexample showing the limitations of mesh-based sharpness guarantees.

The paper addresses a meaningful computational problem in causal inference, with a clear distinction between target preservation, validity, and sharpness.

Critical weaknesses

The core construction builds on existing barycentric stochastic quantization. While the causal lifting interpretation is useful, several theoretical results follow rather directly from linearity, stochasticity, and set inclusion. Hence, I suppose the novelty of contributions is limited.

The sharpness bounds depend on optimal dual solutions of the original problem, which largely defeats the purpose of compression. As a result, the method guarantees validity but does not provide a practical way to determine how much sharpness has been lost.

The improvement over simple geometric refinement is relatively small in most reported settings. E.g., median width inflation decreases only from 2.13% to 1.99% in nonadversarial experiments.

Questions for authors

The authors should prioritize address the weaknesses in rebuttal.

3 comments

We thank the reviewer for recognizing the explicit theoretical constructions, the sharpness counterexample, and the distinction between target preservation, validity, and sharpness. We revised the manuscript and added targeted experiments addressing all three principal concerns.

1. Novelty relative to barycentric stochastic quantization.

Our construction uses established mean-preserving interpolation. The additional theoretical contribution is a finite-IV causal feasible-model transformation.

Theorem 1 maps each original feasible latent response-type distribution to a compressed distribution while preserving its designated causal effect, observation equations, and admissible treatment-response patterns. This proves original-to-compressed identified-set inclusion for the complete set of compatible causal effects.

Theorem 3 characterizes preservation requirements within the shared coordinatewise product-kernel class. Theorem 5 establishes a sharpness limitation that motivates the certified refinement procedure.

We revised Sections 1–3 to explain these results as one coherent causal representation guarantee and to distinguish the mathematical contribution from the underlying interpolation primitive.

2. Practical certification and dependence on original optimal dual solutions.

Your observation about the conditional sharpness-loss bounds is correct: the numerical sensitivity estimates use optimal dual coefficients of the original IV problem.

We now distinguish numerical sharpness-loss estimation from certification of the causal conclusion itself.

The revised Section 4 includes an explicit decision-stopping corollary and short proof.

Let [L,U] be the original sharp interval and [a,b] a certified compressed outer interval. Theorem 1 guarantees

a ≤ L ≤ U ≤ b.

Consequently, if a > τ, every original compatible causal effect exceeds τ; if b < τ, every such effect is below τ.

For τ=0, this provides a certified ATE-sign decision. The stopping rule requires the current surrogate outer interval and threshold, without solving the original LP or obtaining its optimal dual coefficients.

This gives an operational guarantee for a practically relevant task. Numerical certification of the remaining sharpness gap remains a distinct objective addressed by the conditional dual sensitivity analysis.

3. Are adaptive improvements meaningful?

We retain the original nonadversarial experiments, including the modest median width-inflation improvement from 2.13% to 1.99%.

To investigate instances where support allocation substantially affects the result, we added a targeted hard family derived from the sharpness counterexample.

At n=48, M=8, the numbers of certified signs are:

  • Geometric: 12/80
  • Mass-weighted: 72/80
  • Displacement-guided: 80/80

At n=96, M=8, the corresponding results are 6/144, 122/144, and 114/144. At M=16, both mass-weighted and displacement-guided refinement certify all 144 identifiable signs.

These experiments demonstrate that the choice of refinement policy can materially affect the resolution needed to establish a causal decision. They also show that displacement guidance is not uniformly superior: mass weighting performs better at n=96, M=8.

4. End-to-end computational value.

We additionally evaluated complete time to certified decision against strong original-IV LP baselines, including sign-first algorithms. The n=48 experiment uses 80 predefined populations and three fresh-process timing repetitions per case, with all algorithmic refinement overhead included. Mass-weighted refinement certifies 240/240 repeated decisions by M=16, with a median paired exact/compressed runtime ratio of 3.199× on successful certifications.

The revised Section 5 and Appendix C provide the full experimental design, certification outcomes, timing methodology, and limitations of this constructed stress-test family.

Together, the revisions make the causal representation contribution more precise, establish a useful decision certificate independent of the original optimal dual during execution, and directly demonstrate when adaptive refinement matters. We hope these additions provide stronger grounds for reconsidering the recommendation.

Reviewer_2ne5OpenReview ↗

Response to rebuttal

Dear authors,

I remain concerned about the novelty of this work, so I will keep my assessment unchanged. Thank you for your effort.

Best.

Thank you for recognizing the rigorous construction, the sharpness counterexample, and the distinction between target preservation, validity, and sharpness. We have substantially strengthened the third version along precisely these lines.

1. Theoretical contribution and novelty boundary. We distinguish the established barycentric interpolation primitive from the causal model-level guarantee of Theorem 1. Theorem 3 characterizes necessity within the shared product-kernel class, while Theorem 5 proves that mesh size alone cannot uniformly control sharpness loss. Beyond the original affine setting, the new Proposition prop:nonlinear (main Section 3; Appendix A) proves preservation of mixed cross-world moments and polynomial restrictions, including an explicit nonconvex zero-covariance IV example. This extension is intentionally scoped, not presented as a universal nonlinear causal abstraction theorem.

2. We agree with your original-dual objection. The conditional bounds on numerical sharpness loss require optimal dual solutions of the original problem; they are not inexpensive surrogate-only numerical gap estimates. We have made this limitation explicit in Section 4 and Appendix app:dual.

Our practical guarantee answers a different question. If a compressed certified outer interval is [a,b][a,b], then identified-set inclusion implies

a≤Loriginal≤Uoriginal≤b.a\le L_{\mathrm{original}}\le U_{\mathrm{original}}\le b.

Consequently, a>τa>\tau or b<τb<\tau certifies the corresponding threshold conclusion for every original-compatible causal effect. This decision certificate requires neither original optimal dual solutions nor original sharp endpoints during execution (Section 4; Proposition prop:outercomposition; Algorithm 1). We do not conflate decision certification with estimating the remaining sharpness gap.

3. Adaptive refinement makes an important difference in the challenging regime. We agree that ordinary-family improvements are modest: at M=32M=32, median width inflation decreases only from 2.13% (geometric) to 1.99% (dual-guided). We therefore added a counterexample-derived hard family emphasizing actual sign-certification success.

At n=48,M=8n=48,M=8, geometric/mass/displacement refinement certifies 12/80, 72/80, and 80/80 identifiable signs. At n=96,M=8n=96,M=8, the corresponding counts are 6/144, 122/144, and 114/144. Thus no adaptive policy universally dominates; nevertheless, refinement choice can decisively affect whether a valid original-target conclusion is reached.

A separate fresh-process comparison against strong exact HiGHS sign-first baselines measures the complete refinement procedure. For mass-weighted compression at n=48n=48, it reports 240/240 successful repeated decisions and a 3.199× median paired runtime improvement, conditional on successful certification (Section 5; Appendix app:decision-benchmark).

These revisions acknowledge the dual-dependence limitation while demonstrating a separate, operationally useful certificate and substantially more informative evidence on adaptive refinement.

We respectfully invite you to reassess the stronger theoretical scope and decision-oriented experiments in the third version.

Reviewer_XSr6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a certified stochastic compression framework for scalable partial identification in an instrumental variable model. The key idea is to reduce the outcome support using a mean-preserving stochastic mapping while preserving the causal estimand. The authors established theoretical guarantees and proposed an anytime algorithm that tightens the bounds.

Critical strengths
  • The paper rigorously developed the theoretical guarantee using sharpness theory and anytime certification. I find the mean-preserving reduction idea interesting.
Critical weaknesses
  • Although the experiments show substantial improvement in runtime, it is not clear that this solves an important problem in partial identification. The exact bound for IV can be solved efficiently with LP. The exact LP baselines usually take only seconds on synthetic data for a large number of states and less than one second on real-world data. I suppose this could be useful if one wants to estimate CATE with high-dimensional covariates, but that would impose a computational challenge for estimating the distribution. The authors should clarify the practical aspects of the proposed method.

  • The positioning of the paper is not very clear. The major theoretical guarantee seems to follow relatively straightforwardly from existing work on mean-preserving quantization, given the linear constraint and objective for the IV graph. Extending to more complex causal graphs with nonlinear constraints or objectives would be more challenging and interesting, but the paper does not address these settings.

Questions for authors
  • Could the authors please clarify the theoretical challenges the proposed framework solves, given the existing mean-preserving quantization methods?

  • Could the authors provide some examples of practical scenarios in which the speedup improvement can be critical and justify the loss of sharpness?

3 comments

We thank the reviewer for recognizing the theoretical guarantees and interest of mean-preserving reduction. We have revised the manuscript to address your two main questions: the theoretical challenge beyond existing quantization and the practical significance of reducing computational cost.

1. What is the additional theoretical contribution beyond mean-preserving quantization?

We build on established barycentric stochastic interpolation. The central contribution is its integration into a certified finite-IV causal representation.

Preserving the mean of an observed variable is a numerical property. To establish valid bounds for the original causal question, we must additionally preserve the relationship between observed outcomes, latent potential outcomes, treatment-response restrictions, and the designated causal target.

Theorem 1 constructs an explicit latent response-type transformation satisfying these requirements. It maps every original feasible causal model to a feasible compressed model with the same target value. Consequently, every causal effect compatible with the original IV model remains compatible after compression.

This establishes original-to-compressed identified-set inclusion, allowing the smaller causal optimization problem to yield valid bounds for the original target. Theorem 3 characterizes the preservation conditions within the shared product-kernel class.

We have clarified this distinction throughout Sections 1–3 and strengthened the related-work comparison with quantization, causal abstraction, and existing IV optimization. The current guarantees cover finite-IV response-type models with designated affine mean-based targets. More general nonlinear causal constraints would require corresponding feasible-model preservation results.

2. In which practical scenarios can the acceleration justify reduced sharpness?

We agree that an individual original IV LP can be inexpensive, as illustrated by the JTPA application.

The revised manuscript now makes a second computational objective explicit: time to a certified causal decision, rather than only time to both sharp numerical endpoints.

For example, consider repeated analyses of whether a treatment effect is positive across outcome resolutions, robustness specifications, or populations. If the compressed interval lies entirely above zero, every causal effect compatible with the original IV model is positive. The researcher can stop without recovering the original sharp interval.

The justification for reduced sharpness is precise: once the interval certifies the desired threshold conclusion, its additional width cannot change that conclusion. If the interval remains inconclusive, refinement continues.

The new corollary in Section 4 proves this stopping condition directly from Theorem 1. Algorithm 1 implements the corresponding anytime procedure.

3. New evidence against strong exact baselines.

We added a targeted family of difficult IV populations and measured the complete cost of obtaining a certified sign conclusion.

At n=48 and retained support M=8, geometric, mass-weighted, and displacement-guided refinement certify 12/80, 72/80, and 80/80 identifiable ATE signs, respectively.

We also compare against original-IV LP implementations, including sign-first methods that can terminate after solving a single endpoint. In 80 population designs with three serial, single-thread timing repetitions, mass-weighted refinement certifies 240/240 repeated decisions by M=16. Its median compressed time is 0.0183 seconds, with a 3.199× median paired runtime ratio relative to the fastest certified exact sign-first baseline on successfully certified cases.

The larger-support benchmark separately demonstrates substantial computational acceleration for numerical bound computation, while JTPA establishes compatibility with an empirical IV application.

Covariate-stratified analyses could involve repeated IV queries, although estimating high-dimensional conditional distributions introduces additional statistical challenges and is outside our experimental evaluation.

We hope the clearer causal novelty statement, explicit decision-level guarantee, and stronger matched computational comparisons address your concerns about theoretical significance and practical usefulness.

Thanks for the detailed responses. I appreciate the clarification on the contribution and motivation of the proposed framework. However, my main concerns are not fully resolved.

I understand that the proposed framework establishes identified-set inclusion, which goes beyond simply preserving the mean of the observed variable. However, given the existing work on mean-preserving quantization and the linear structure of the IV model, the contribution still seems incremental and does not appear to be a sufficiently strong methodological or theoretical contribution.

Also, I'm not fully convinced about the practical significance of the proposed method. Although the new experiments demonstrate speedups, it is unclear in which scenarios these would be critical.

Overall, my main concerns regarding the significance of the contribution and practical necessity of the proposed method remain. I believe the paper would benefit if extended to cases where the sharp bounds cannot be solved with linear programming; such an extension would require nontrivial theoretical development and would also bring more practical value to the framework.

Thank you for identifying the mean-preserving reduction idea as interesting and requesting stronger evidence of theoretical novelty and practical necessity. The third manuscript version now directly addresses your suggestion to move beyond linear causal objectives and constraints.

1. Novelty beyond established quantization. We agree that mean-preserving barycentric quantization is established. Our contribution is not the interpolation identity itself, but its causal representation certificate: Theorem 1 maps every original feasible latent IV model to a compressed model preserving its observed-data equations, treatment-response restrictions, and designated target. This proves one-sided identified-set inclusion, rather than merely preserving an outcome expectation. Theorem 3 characterizes necessity within the stated product-kernel class.

2. New nonlinear theorem and nonconvex example. Main Section 3 and Appendix A, Proposition prop:nonlinear, now establish preservation of cross-world mixed moments and polynomial functions and restrictions built from them. In particular, the lift preserves E[Y0Y1]\mathbb E[Y_0Y_1] and Cov⁡(Y0,Y1)\operatorname{Cov}(Y_0,Y_1). We provide an explicit fixed-observation example in which imposing zero covariance creates a nonconvex IV feasible set, while original-to-compressed identified-set inclusion still holds. The theorem precisely delineates the preserved-moment class; we do not claim to solve arbitrary nonlinear causal graphs. Section 2 explicitly distinguishes this result from dual quantization, prior potential-outcome identification, and Duarte et al.'s existing polynomial causal optimization.

3. Computational significance beyond fast individual LPs. We agree that a single sharp IV LP may already be inexpensive, as in our JTPA example. We therefore evaluate whether compression can establish a scientifically sufficient original-target decision before recovering numerical sharpness. On a deliberately challenging n=48n=48 family, geometric, mass-weighted, and displacement-guided refinement certify 12/80, 72/80, and 80/80 identifiable ATE signs at M=8M=8. A separate matched sign-first benchmark achieves 240/240 successful repeated decisions, with a 3.199× median paired runtime improvement for mass-weighted compression among successful decisions (Section 5; Appendix app:decision-benchmark).

4. New nonlinear optimization experiments. Appendix app:nonlinear-exp reports six matched n=64,M=8n=64,M=8 covariance instances. Both representations certify the same exploratory threshold in all six, with 1.52× median paired acceleration under direct SCIP and 70.77× under HiGHS spatial McCormick LP relaxations. The solver formulations and timing protocols are stated separately. In one n=128n=128 pilot, compressed SCIP obtains a conclusive upper bound within the budget while original SCIP remains inconclusive; original-resolution HiGHS also provides a conclusive bound. We do not present these as AutoBounds front-end timings or universal nonlinear speedups.

The new nonlinear preservation theorem, explicit nonconvexity, and matched computational results substantially strengthen the theoretical and practical contributions beyond the initial linear-IV presentation.

We respectfully ask you to evaluate these additions in the third version and reconsider whether the contribution remains incremental.

Reviewer_n8AxOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

Computing sharp bounds on the ATE in a discrete observations IV setting in the potential outcomes framework becomes computationally challenging as the outcome set cardinality grows, specifically sharp bounds problem is an LP with dimension polynomial in outcome set cardinality.

The main theorem is a population-level statement. There's a finite-sample result given empirical counts leveraging as Bonferroni / union bound type simultaneous confidence interval.

The paper proposes an outcome coarsening device that preserves validity, that is outputs a set that contains the sharp identified set, while greatly alleviating computational burden.

Critical strengths

Well motivated and grounded in practical use cases: cases where computing sharp bound on ATE are plentiful, yet practitioners often don't because it's computationally daunting.

Paper appears correct throughout.

The formulation is quite elegant and simple. Numerics are convincing.

Critical weaknesses

I find it extremely hard to read. A lot of non-canonical concepts are not defined before the first time they appear in the main text. Readers can't be expected to know what the authors mean by "a column-stochastic outcome map" - a column-stochastic matrix is a standard concept, an outcome map isn't quite and much less so a column-stochastic outcome map. Also, "support uniform mesh-only rate" in the abstract sounds very obscure to me.

Narrow the scope and avoid 5 contributions: condense your contributions to one population-level computational punchline, one inference theorem and and supporting simulation and real data experiments.

Conservativeness of the Bonferroni based procedure isn't very appealing but you clearly acknowledge it. But it's also not clear in what sense: is it in expected interval length in an appropriate minimax sense? A formal theorem would be helpful.

I recommend weak rejection based mainly on that as I find the core idea quite clever and elegant and the problem of high practical importance. I have no issue with the rest of the paper. I'm open to read an updated manuscript and change my rating.

Questions for authors

By how much is the Bonferroni device conservative in expected interval length. Wouldn't there be easy tightenings?

3 comments

We thank the reviewer for recognizing the elegance, correctness, and practical importance of the core idea. Your comments were especially helpful in restructuring the manuscript to make its main guarantees easier to understand.

1. Clarity, terminology, and contribution structure. We substantially revised the abstract, Introduction (Section 1), and mathematical exposition (Sections 3–4). We now introduce stochastic outcome maps through a concrete numerical example before defining their mathematical properties. We also explain response types, the product lift, mean preservation, and outer bounds in plain language alongside their formal definitions.

Following your recommendation, the Introduction now presents two main contributions: (i) certified population-level causal compression and its computational consequences, and (ii) nonasymptotic finite-sample inference. The necessity characterization, sharpness counterexample, and refinement guarantees support this structure. The main text also explains the central proof mechanism, with complete technical proofs retained in Appendix A.

2. How conservative is Bonferroni in expected interval length? We added a direct paired comparison on 24 synthetic IV populations, with 50 independent sampling repetitions per population at each sample size.

At nominal 95% confidence, the mean robust ATE interval widths for Bonferroni–Clopper–Pearson (CP) versus Dvoretzky–Kiefer–Wolfowitz (DKW) inference are:

Sample sizeCPDKWMean-width reduction
5001.08810.615443.4%
1,0000.87230.513941.1%
5,0000.51200.370027.7%

Appendix B reports paired Monte Carlo uncertainty. All methods achieved empirical coverage of 100% in these simulations; their formal simultaneous-coverage guarantees are at least 95%.

3. Formal tightening and coverage. We also added a confidence-region pushforward proposition in Appendix B. It proves that a simultaneously valid confidence region on the original observed law retains its population identified-set coverage guarantee after admissible compression, including data-dependent selection of refinement knots.

We instantiate this result using DKW cumulative-probability bands and a valid CP–DKW hybrid. The construction operates on original observed categorical counts, preserving the appropriate sampling model. A separate compressed-DKW experiment demonstrates small additional interval width at retained support M=8.

These results answer the expected-length question quantitatively on the evaluated populations and provide formal coverage guarantees for tighter alternatives. A minimax expected-length characterization remains a separate statistical question.

We hope the substantially clearer presentation, focused contribution structure, new formal inference result, and direct conservatism measurements address the principal concerns motivating your initial recommendation. We would welcome your reassessment of the revised manuscript.

Thank you for recognizing the elegance, apparent correctness, and practical importance of our approach. We have substantially revised the third manuscript version to address the presentation and statistical concerns that motivated your recommendation.

1. Clearer exposition and exactly two contributions. Sections 1 and 3 now explain response types, the probabilistic outcome map, column stochasticity, mean preservation, the product lift, and outer validity in accessible language before developing the formal theory. We reorganized the paper around two principal contributions: (i) population-level certified causal compression and its decision-stopping consequence (Theorem 1 and Section 4), and (ii) finite-sample outer inference (Theorem 8). Necessity, sharpness limitations, and nonlinear preservation are now supporting results rather than competing contribution claims.

2. Quantifying Bonferroni conservatism. We added a paired study over 24 synthetic IV populations, with 1,200 simulated intervals per sample size. At N=500N=500, mean robust ATE interval width decreases from 1.0881 under Clopper–Pearson/Bonferroni to 0.6154 under DKW, a 43.4% reduction. The corresponding reductions at N=1,000N=1,000 and N=5,000N=5,000 are approximately 41.1% and 27.7%. Section 5 and Appendix B (Table tab:dkw-width-comparison) report the comparisons and Monte Carlo uncertainty.

3. Formal validity of tighter inference. Appendix B establishes a confidence-region pushforward result: simultaneous DKW regions and CP–DKW hybrid regions retain their finite-sample coverage after admissible outcome compression, even with data-dependent refinement. This provides a mathematically valid tightening, not merely a numerical heuristic. We distinguish this coverage theorem from the empirical expected-width comparisons and do not claim a minimax expected-length result.

We believe these revisions directly address your principal objections. We respectfully invite you to read and reconsider the substantially clearer third version.

Dear authors, Will read the updated draft during reviewer discussion period.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.