acceptodds
Under review as a conference paper at ICLR 2027

Sharp Certificates for Worst-Case Controlled Cost Sensitivity

Abstract

We study exact identification and finite-sample certification of the worst-case controlled contrast . The access contract is explicit: fresh bounded joint interventions fix every nuisance coordinate, and a valid global Lipschitz constant is supplied externally. For any feasible collection of simultaneous grid-value intervals, McShane–Whitney envelopes give the exact identified interval for ; at zero trace its unavoidable geometric width is . Thresholding its confidence refinement yields family-wise valid sensitive, insensitive, or unresolved reports. For adaptive placement we prove a KL-weighted cap- packing lower bound and a doubling-tree upper bound. On , a converse valid for every -Lipschitz insensitive base, without the lower comparison's strict-slope or interiority assumptions, bounds the packing functional by . The reverse bound, and hence the two-sided comparison, still require an -Lipschitz base and interior contrasts; corrected constants improve by factors 8, 16, and 32 for but remain loose. On the declared synthetic bank, Lipschitz placement remains the strongest equal-cap comparator (1,155 versus 1,037 stagewise decisions at 1,048,576 pairs), and understating causes false-low reports. A retained execution is reported only as an aggregate theorem audit, not a competitive evaluation. The result certifies controlled sensitivity under this oracle contract; it does not learn , identify causal parents, improve a controller, or establish policy safety. \

Primary area:
causal reasoning
Keywords:
simultaneous confidence, controlled interventions, Lipschitz functions, selective reporting

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.80 over 5 reviews, above 27% of the others.
Reviewer_GNfmOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper studies worst-case contrast, i.e., whether MiM_i is above or below a threshold. It provides SENSITIVE, INSENSITIVE, or UNRESOLVED reports and confidence bounds under the stated setting and assumptions. Additionally, it analyzes adaptive sampling/placement, presenting a cap-packing lower bound and a doubling-tree upper bound. Performance evaluation is reported using synthetic experiments.

Critical strengths

The paper derives an adaptive query-complexity lower bound and a doubling-tree upper bound. These bounds appear to be meaningful contributions, though the constants are loose.

The paper characterizes the full range of worst-case effects consistent with those bounds, considering a fixed set of confidence intervals at sampled settings and a known Lipschitz bound.

Critical weaknesses

The reviewer found the paper quite difficult to follow and comprehend, and it is difficult to identify the main idea or concept, making the contributions difficult to assess and validate.

One concern of the reviewer regards the practical relevance of the presented analysis, i.e. the assumptions seem quite restrictive, which limits applicability. For instance, the authors assume an externally validated global Lipschitz constant and require fresh joint interventions fixing every nuisance coordinate.

The novelty appears to be limited in the fixed-grid results. The confidence bounds combine standard arguments (concentration and Lipschitz covering); the exact envelope result applies classical McShane–Whitney extensions. The empirical evaluation is synthetic and relatively limited, and the paper also does not show a practical advantage over existing Lipschitz methods (the proposed method is not shown to outperform Lipschitz placement, at least in the larger reported budgets).

Questions for authors

Q1: What specific guarantee in Theorems 4C–4D is not covered by existing certified Lipschitz optimization methods?

Q2: Can you give a realistic setting where the required joint interventions and externally validated Lipschitz constant are available?

Reviewer_cGfdOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This work talks about worst-case controlled contrast

Mi=sup⁡z∣Δi(z)∣M_i=\sup_z|\Delta_i(z)|

under bounded paired interventions and a known global Lipschitz constant. This work tries to combines pointwise confidence intervals with Lipschitz envelopes and respond as SENSITIVE, INSENSITIVE, or UNRESOLVED.

The work also points out exact set of possible supremum values for one fixed interval-valued grid trace. For adaptive querying, it provides a cap-packing lower bound and a doubling-tree upper bound. The experiments compare certified allocation rules on synthetic functions.

Critical strengths
  1. The paper gives the feasibility condition and the exact minimum and maximum values of sup⁡z∣f(z)∣\sup_z|f(z)| that agree with the observed intervals. It also proves that every value between the two endpoints is attainable.

  2. The paper does not call controlled sensitivity causal-parent discovery or policy safety. The adaptive proof also avoids an optional stopping error by using fresh predictable batches and a finite union over predeclared node stage events.

Critical weaknesses
  1. McShane-Whitney extensions, level-set estimation, certified Lipschitz optimization, packing lower bounds, and margin-based instance complexity already provide much of the surrounding framework. The fixed-trace result is useful. However, the remaining contribution is mainly a careful specialization to this sensitivity target.

  2. No external dataset satisfies the required combination of full nuisance intervention and an independently valid global LL. The experiments are fully synthetic. Even there, Lipschitz-envelope placement resolves 1,155 candidates at the largest equal budget, while the proposed stagewise method resolves 1,037. It is also not ICLR quality paper .

Questions for authors
  1. What is the strongest theorem here that cannot be recovered by combining the classical Lipschitz envelopes with an certified optimization or level-set algorithm?

  2. Please clarify the statement that kl/γ2=57.06\mathrm{kl}/\gamma^2=57.06 at τ+γ=.99\tau+\gamma=.99. Keeping the earlier γ=.06\gamma=.06 gives approximately 30.16.

  3. Is there a realistic domain where both the required intervention oracle and an independently valid global Lipschitz constant are available?

Reviewer_8Z3VOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

The paper asks how to certify whether a controlled effect stays below a threshold across all reference settings. It assumes that a valid Lipschitz constant is given. Under this assumption, it gives the exact range of worst-case effects that fits the interval estimates at the sampled points. It also gives guarantees for adaptive sampling. The results are checked with synthetic experiments and numerical audits.

Critical strengths
  1. The exact interval characterization is useful. It shows what the collected data actually establish. It also explains why sampling the same points many times can still leave uncertainty between them.

  2. The cap-packing lower bound gives a meaningful picture of how hard the sampling problem is. The paper also states the assumptions and limits of its guarantees carefully.

Critical weaknesses
  1. Matching the upper and lower bounds needs several strong assumptions. These include a much smoother base function, contrasts away from the range boundaries, and specific conditions on the sampling schedule. The paper states them clearly, but they limit where the matching result applies.

  2. The complexity upper bounds are still quite loose in the numerical audit. The smallest complete upper bound in Table 9 is 1,401,633, which is larger than the deterministic budget cap of 516,096. This does not break the theory. But it means the bounds say little about the sampling cost on the reported instances.

  3. The experiments do not clearly show that the main new components help. Table 2 evaluates the budget-recycling version of Theorem 4A. Theorem 4D is tested separately and is not compared with the same baselines. The layout makes this worse. The main text is only six pages, but the key experimental tables are in the appendix, and no figure gives an overview of the method. Moving the main comparisons and a small worked example into the main text would make the contributions much easier to judge.

Questions for authors
  1. Is there an example that satisfies your assumptions where the complete complexity upper bound is smaller than the deterministic budget cap?

  2. Can you give a simple example where the exact interval in Theorem 2A reaches a clear decision (sensitive or insensitive), but the standard bound leaves the case unresolved?

  3. Can you compare the doubling-batch algorithm in Theorem 4D directly with the baselines in Table 2, using the same budget and the same error level?

Reviewer_p3hmOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies finite-sample certification of worst-case controlled cost sensitivity under known Lipschitz constraints. It derives sharp identified intervals, family-wise valid sensitivity decisions, and adaptive query-complexity bounds. Synthetic experiments evaluate the proposed certificates and sampling strategies, including their sensitivity to misspecified Lipschitz constants.

Critical strengths
  1. Strong theoretical formulation. The distinction between detecting sensitivity and certifying insensitivity is well motivated. I particularly appreciate the sharp fixed-trace identification result and its finite-sample guarantees.

  2. Interesting adaptive analysis. The cap-packing lower bound and doubling-tree upper bound provide useful insights into instance-dependent sample complexity. The paper also carefully discusses the limitations of its guarantees.

Critical weaknesses
  1. Strong assumptions. The guarantees require a known valid Lipschitz constant and full control over joint interventions, which may be difficult to satisfy in practical settings.

  2. Limited empirical validation. All experiments are synthetic, and the proposed stagewise approach does not consistently outperform existing Lipschitz-based methods. This leaves its practical advantage somewhat unclear.

Questions for authors
  1. How could a valid Lipschitz constant be obtained in practice? How sensitive is the method to using an overly conservative estimate?

  2. Under what conditions would the proposed adaptive approach outperform standard Lipschitz-envelope placement, given the results in Appendix Table 2?

Reviewer_HGb6OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies the effect of individual features on a predefined cost function, quantified through conditional causal effects under "interventions". The authors derive high confidence bounds for estimating this quantity. Their approach relies on reducing the sup over the entire Euclidean space to a finite, predefined grid, using a Lipchitz assumption on the quantity of interest. However, this assumption is neither sufficiently justified nor in general guaranteed to hold for conditional causal effects under interventions. Therefore, the validity and applicability of the proposed guarantees remain unclear.

Critical strengths

Authors provide a theoretical analysis and derives high-confidence bounds by applying and adapting results from [1] and the attempt to provide finitesample guarantees for evaluating feature effects is potentially interesting and relevant.

Critical weaknesses

I found the paper particularly challenging to read due to issues with its presentation, lack of motivation, and technical clarity.

Absent motivation and Recent Related Literature. The paper lacks motivation for the problem. Mainly, the authors don't explain why the proposed property of conditional causal effect is relevant in that context. This also results in a limited discussion of related works. Furthermore, the paper does not appear to engage sufficiently with recent literature in this area.

Insufficient Experimental details. The experimental section is absent in the main paper and does not provide sufficient detail to understand the experimental setup or interpret the reported results. Despite the title of Section 4, the section does not appear to be as a comprehensive experimental evaluation or a conclusion. The conclusion is instead given in Appendix C. Moreover, the paper does not provide pseudocode or a description of the proposed evaluation procedure. This makes it difficult to understand precisely what is being evaluated and how the proposed method operates in practice.

Lack of clarity in proof writing notations. Several mathematical objects and notations are introduced without explanation. This issue is already apparent in the abstract, where mathematical expressions and terminology are introduced before the underlying concepts are properly motivated or defined. For example, the sentence: "Thresholding its confidence refinement yields family wise valid SENSITIVE, INSENSITIVE, or UNRESOLVED reports." introduces terminology that is not explained until Section 2. This makes the abstract unnecessarily difficult to follow. More generally, the paper frequently introduces equations and notations without sufficiently explaining their meaning or their role in the argument. Several definitions and statements also appear ambiguous or inconsistent, as detailed below.

Technical contribution and justification of assumptions. Some of the theoretical results appear to be direct extensions or applications of existing results, such as Theorem 1, which seems to follow directly from [1]. Other results rely on assumptions that are not adequately justified. In particular, the Lipschitz assumption underlying the finite-grid approximation is central to the proposed guarantees, yet its validity for the conditional causal effect under consideration is false.

Overall, I believe the paper requires substantial revision, both in its technical exposition and in its motivation and experimental evaluation, before it would be ready for consideration at a selective conference such as ICLR.

Minor comments.

  • The finite grid is defined over the space X−iX_{-i}. This should be stated explicitly and consistently throughout the paper.

  • Line 117: PC=Pa(C)∩XP_C = P_a(C) \cap X. The quantity Pa(C)P_a(C) does not appear to be defined.

  • Line 132: The expression " nikn_{ik} independent pairs produce" should presumably read " zikz_{ik} independent pairs produce."

  • Line 132: In EDikr=Δi(zik)\mathbb{E}D_{ikr} = \Delta_i(z_{ik}), the left-hand side depends on rr, whereas the right-hand side does not. Could the authors clarify this discrepancy and justify the equality?

  • Line 135: The statement "Independence is required across replicate pairs, not across grid points or candidates" is unclear. Could the authors clarify precisely which random variables are assumed to be independent and how this assumption is used?

  • Line 195: The authors refer to the distance between a number and a set. This quantity has not been defined.

  • Line 1556: The statement "Necessary, per candidate. Planting one tent of height $2\gamma $ and slope ..." is hard to understand for me. Please clarify the intended meaning.

  • Line 1620: sentences ending with "py" files instruction appears unusual and is not sufficiently explained.

Questions for authors

Definition and motivation of the conditional causal effect. I am still unsure about the precise definition of the conditional causal effect used in the paper. As I understand it, the function allows zz to vary, rather than fixed. This raises a fundamental question about the motivation behind the proposed quantity: why isn't zz fixed when measuring the effect of intervening on a particular feature ii? Could the authors explain the motivation for this choice and its implications for the interpretation of the resulting quantity?

Definition of the one-step cost function. What exactly is meant by a one-step cost function? Where does the temporal dimension enter the problem ?

Clarification of Theorem 2 and the exact grid trace. In Theorem 2 (Line 182), how is the exact grid trace defined? Previously, Δ\Delta was defined with respect to a fixed feature and indexed by ii. Here the notation seems to refer to a different quantity which has not been introduced.

If index ii is intended, then same setting Δi\Delta_i to zero would imply that the grid is invariant with respect to feature ii, meaning for any zz in the grid, flipping its ii-th feature would result in another point belonging to the grid. This would amount to an invariance property with respect to feature ii. Is this the intended interpretation? A similar issue arises in Theorem 2a, where GG is no longer indexed by ii. The authors do not explain how the problem is being reformulated and the resulting quantity is not defined. More generally, Theorem 2 requires more precise statement of what is meant by an "exact grid trace." In particular, the value of rikr_{ik} appears to be strictly positive, so how can the exact interval be obtained if rikr_{ik} can never be zero?

Clarification of the evaluation protocol and the role of out-of-distribution samples. In the absence of pseudocode describing the interactive protocol, it is difficult to understand how the algorithm constructs its queries. In particular, when evaluating the conditional causal effect, does the algorithm explicitly construct queries that account for potential confounding? My understanding is that the algorithm does not rely on out of distribution samples. However, this point deserves an explicit discussion, particularly regarding how the proposed guarantees relate to the distribution of queried samples and whether the evaluation procedure requires interventions that may generate out of distribution inputs.

[1] E. J. McShane. Extension of range of functions. Bulletin of the American Mathematical Society, 40(12):837–842, 1934. doi: 10.1090/S0002-9904-1934-05978-0.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.