acceptodds
Under review as a conference paper at ICLR 2027

When Does a Remaining Privacy Budget Preserve Every Admission Decision?

Abstract

A reusable privacy budget should preserve the choice of the next release. We ask when one experiment represents exactly the requests admitted by the original spent–cap comparison after independent composition. Every cap with a nonempty feasible set has a universal residual if and only if the spend has at most one finite privacy-loss value, with complete reveal as the degenerate case. The key is how composition acts on request profiles. One value rescales the whole profile; several values average different rescalings and can conceal a crossing. We realize this obstruction with valid, bounded-likelihood-ratio experiments for every finite-support multi-atom spend, with ordered compositions and a cap equal to one of those products. Weakly continuous composition semigroups with finite privacy-loss support at each time have the one-value structure, making parameter subtraction exact for arbitrary requests. A concrete Gaussian spent–cap pair has no universal residual and admits a randomized-response channel that its Gaussian-family residual excludes. An exact finite example visualizes the obstruction, and a model-selection illustration separates the additional channel accuracy from its downstream utility. The results identify when a single remaining-budget experiment preserves the choice of the next mechanism.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
differential privacy, Blackwell order, privacy composition, privacy residuals, statistical experiments

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_Uj6POpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

The paper considers the question whether we can characterize the remaining budget in fully adaptive f-DP accounting via a pair of distributions (P,Q)(P,Q) such that all possible choices of mechanisms (possible meaning all mechanisms for which the total trade-off function does not cross a predefined threshold trade-off function) lead to total trade-off function that is above the trade-off function we'd get when we compose the trade-off function determined by the pair (P,Q)(P,Q). This is a sort of subproblem one encounters when trying to carry out fully adaptive privacy accounting with trade-off functions or with privacy loss distributions and privacy profiles. The answer is negative: only in case of point mass and certain "total-failure probability" mechanisms this is possible. The background for this problem is fully adaptive accounting: Regehr et al. and Long et al. (references in the paper) have recently shown that in general it is not possible to do fully adaptive accounting accurately with so called trade-off functions or privacy loss distributions. It is possible with Rényi DP (Zrnic and Feldman, 2021), and it is an interesting question whether something more accurate is possible since in the non-fully adaptive case we can do more accurate accounting than RDP, via privacy loss distributions.

Critical strengths
  • Interesting characterization for this sub problem of fully adaptive DP accounting.
Critical weaknesses
  • The motivation is unclear. Suppose we had a meaningful positive answer: there is a non-trivial class of mechanisms for which we can characterize that remaining budget via single trade-off function or single pair of distributions. What could we do with this characterization? I think we could not somehow 'subtract' the trade-off function from the 'budget trade-off function' as in case of GDP, for example. There is that Torgersen's product theorem given which considers non-adaptive compositions, but I don't see how it could be applied here. This leads me to the next point:

  • I think the setting the paper considers is ultimately non-adaptive. I don't exactly see how the results could be applied in the fully adaptive case.

  • The writing could be improved. I understood the main results of the paper but to be honest, I couldn't really understand some parts like the abstract. Also elsewhere there are sentences like: "Feasibility supplies safety; greatestness supplies completeness." which is I cannot understand.

  • I think one weakness is that the problem has been in some sense already solved by Regehr et al. and Long et al. (references in the paper). They both show that in general it is not possible to do fully adaptive accounting accurately with so called trade-off functions or privacy loss distributions. That suggests that for general-purpose fully adaptive accounting (and for fully adaptive accounting for DP-SGD) that is more accurate than the RDP one the right direction is some sort of approximate accounting. That does not in itself of course exclude accurate fully adaptive accounting in some special cases, but the paper does not seem to identify any such special case.

Questions for authors
  • same as above: Suppose we had a meaningful positive answer: there is a non-trivial class of mechanisms for which we can characterize that remaining budget via single trade-off function or single pair of distributions. What could we do with this characterization? We could not somehow 'subtract' the trade-off function from the 'budget trade-off function' as in case of GDP, for example, and operationalize that for fully adaptive accounting, or?
3 comments

Thank you for your careful review. We appreciate your constructive comments.

We have uploaded a revised version, with red margin notes naming the review point each change addresses.

W1 (motivation), Q1 (use of a positive answer)

An exact one-experiment remaining budget is essentially unavailable, and the revised abstract and introduction state this negative result as the main conclusion. Theorem 1 shows that a spend CC has a greatest feasible request at every feasible cap exactly when its privacy loss, whenever finite, takes a single value. This class is of erasure type: one common output, plus outputs that reveal the hypothesis. Gaussian, Laplace, subsampled Gaussian and randomized-response mechanisms lie outside it, so each has a cap BB at which no single experiment replaces the pair (C,B)(C,B) (new Table 1 and the paragraph after Theorem 1).

Where a positive answer exists, it is used by subtraction, as for GDP; the revised Section 6 opens with this use. For a weakly continuous semigroup with finite privacy-loss support, Theorem 4 and Eq. (2) give

Cs⊗R⪯Ct⟺R⪯Ct−sC_s\otimes R\preceq C_t\quad\Longleftrightarrow\quad R\preceq C_{t-s}

for every binary request RR, including requests outside the family, so Ct−sC_{t-s} alone answers every one-step admission question. The same theorem shows that these semigroups have one finite loss value at each time and are erasure semigroups in the self-dual case. Theorem 4 therefore marks the boundary of the negative result, and the revised Section 6 closes with this statement and a worked erasure example.

Torgersen's theorem enters the negative results as the cancellation step. In Section 5, take the cap B=C⊗R2B=C\otimes R_2. A greatest feasible R⋆R^\star would satisfy R2⪯R⋆R_2\preceq R^\star and C⊗R⋆≃C⊗R2C\otimes R^\star\simeq C\otimes R_2; cancellation gives R⋆≃R2R^\star\simeq R_2, which contradicts the feasibility of R1⪯̸R2R_1\not\preceq R_2.

W2 (non-adaptive setting)

The setting is one step with independent composition C⊗RC\otimes R, as the revised introduction states; fully adaptive filtering is the subject of Regehr et al. and Tran et al. One step is where a remaining budget is used: under GDP, after spend G1G_1 against cap G2G_{\sqrt2} the analyst passes on G1G_1 alone, and the next Gaussian request is compared with it without the original pair. Theorem 1 says exactly when this hand-off loses nothing, and Example 1 shows what it loses otherwise. The direct test admits RR(0.70), while the residual G1G_1 allows at most RR(0.6915), a bit error of 30.854% against 30.000% (Table 3), and no feasible residual admits both G1G_1 and RR(0.70) (Table 2). The loss is thus a property of the spend, present for one independent release.

W3 (writing)

The spend CC bounds past disclosure and the cap BB bounds total disclosure. A request RR is admitted when C⊗R⪯BC\otimes R\preceq B. A remaining budget is one experiment R⋆R^\star such that RR is admitted exactly when R⪯R⋆R\preceq R^\star. "Feasibility supplies safety" means that R⋆R^\star is itself admitted, so every request below it is admitted. "Greatestness supplies completeness" means that every admitted request lies below R⋆R^\star. The revised Section 3 uses these explicit statements, and the revised abstract and introduction define each term at first use.

W4 (relation to prior work)

For each standard mechanism Theorem 1 gives a definite rule (revised Section 1, after Table 1): keep the pair (C,B)(C,B), or export a family-restricted or approximate residual, whose loss Example 1 quantifies. Erasure-type spends, and only these, admit an exact residual at every feasible cap.

The negative direction of Theorem 1 uses the supremum construction, the composition criterion (their Theorem 7, our Proposition 2) and the gap mechanism of Regehr et al., whose Corollary 4 establishes failure under some Gaussian budget. The last paragraph of the revised introduction lists three additions.

First, Theorem 1 decides every fixed spend, with an arbitrary privacy-loss law. The criterion asks whether composition with CC preserves suprema of requests; for each nonconstant finite-loss law, Appendix D.3 builds two requests whose profiles cross at a point that splits the law, so their supremum is not preserved.

Second, Theorem 3 (Section 5) separates equivalence cancellation from order reflection. For every finite-support spend with several finite loss values, it constructs two bounded, mutually absolutely continuous requests whose profiles cross while their products with the spend are ordered at every threshold; the larger product is a cap with no greatest residual.

Third, Theorem 4 (Section 6) classifies weakly continuous semigroups with finite privacy-loss support at every time.

We hope these responses and the revision address your concerns, and we respectfully invite you to reassess the paper in light of them.

Thank you for the reply. Regarding the first point: that semigroup property does hold for the Gaussian mechanism (or GDP mechanisms, when we are carrying out fully adaptive GDP accounting, or other single-parameter family of trade-off functions with ordering), and that can be used for the fully adaptive accounting. However, in the paper, Section 3 considers general mechanisms and the results show impossibility in the general case. I don't see how it could be used for fully adaptive accounting. In particular, I don't see how the equivalence given at the end of page 3 could be used for fully adaptive accounting, since you need to be able to somehow subtract that R∗R^*.

Thank you for the follow-up. The equivalence on page 3 is used by taking R⋆R^\star as the new cap. Suppose R⋆R^\star exists for (C,B)(C,B) and the next request R2R_2 is chosen after seeing the output of CC, subject only to R2⪯R⋆R_2\preceq R^\star. Every such conditional release is then bounded by the same fixed experiment R⋆R^\star, so the adaptive composition theorem of Dong et al. (2022) bounds the whole interaction by C⊗R⋆⪯BC\otimes R^\star\preceq B. The accountant can then discard CC and BB: the next remaining budget is the residual of the issued R2R_2 at the cap R⋆R^\star, again a one-step question. Applying the page-3 equivalence to the request R2⊗RR_2\otimes R shows that this residual coincides with the residual of C⊗R2C\otimes R_2 at BB, and the same argument, conditional on the first output, covers each later step. The subtraction is therefore the residual map (C,B)↦R⋆(C,B)\mapsto R^\star applied repeatedly, and it permits each request to depend on earlier outputs whenever every residual along the way exists. Under GDP this map is the subtraction of squared parameters.

Section 3 concerns general mechanisms, and Theorems 1 and 3 identify when this map is defined: at every cap exactly for erasure-type spends, while Gaussian, Laplace, subsampled Gaussian and randomized-response spends each have caps at which it is undefined. At such a cap the admitted requests have no admitted upper bound, so an accountant that lets the next request depend on earlier outputs has no single experiment to compose against; Regehr et al. (Corollary 4) exhibit an adaptive failure under a Gaussian budget. Our results locate this obstruction already in one step. We will add this paragraph to Section 3 of the final version.

Reviewer_gXyoOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
1
Soundness
3
Read review
Paper summary

This work studies the question in exact (PLD/f-DP based) differential privacy accounting of when a remaining privacy budget can fully capture the admission rule based on exact composition. They consider a specific version of this question in which some queries with a total cost of C have been issued, an overall privacy budget of B is to be maintained, and one would like to capture the one-step admission rule C⊗R⪯BC \otimes R \preceq B in terms of a remaining budget R∗R^*. The authors state that this is equivalent to the question of understanding when the residual class {R:C⊗R⪯B}\{R : C \otimes R \preceq B\} has a maximal element.

The first result states that, for a fixed and already accumulated cost CC, the residual class has a maximal element for an arbitrary choice of budget BB only when CC is degenerate, i.e. its finite support has at most one element.

The second result I cannot parse fully as the exposition is quite unclear but appears to be very similar in spirit to the first result.

The third result consider the restriction to spent costs CC and budgets BB belonging to a well-structured family indexed by a scalar. Further assuming that this family only includes elements of finite support, the authors show a similar result that the residual classes contain maximal elements only when the family consists of degenerate costs.

Critical strengths

None

Critical weaknesses

The writing and framing of this paper and its results is overall very unclear. The central question is not clearly motivated. The phrasing of the results makes them difficult to interpret. It is also hard to understand their significance as there is no intuitive explanation of the implications. As an example, the role of "equivalence cancellation" in the statement of Theorem 3 is hard for me to parse and it is also not clear why it is important to characterize when it occurs. The authors tend to use new terms/phrases without clearly introducing them, e.g. "spend", "cap", or "preserve every feasible choice of the next release". These interruptions make it harder to follow the story.

In addition, as far as I can tell, the paper does not offer substantial new technical contributions. The results appear to be straightforward extensions of the Regehr et al. 2026 work. For example, Theorem 1 can essentially be shown by instantiating the condition in Thm 7 from the cited work with an appropriate example.

Overall, I think this work could be substantially improved by rewriting it to tell a clearer more coherent story throughout. The negative results presented in this work do seem potentially interesting but I think are held back by a lack of convincing narrative behind them.

Questions for authors

Can the authors explain the motivation for the one-step question and what the intended application/interpretation of their results about it should be?

4 comments

Thank you for your thoughtful review. We appreciate your detailed questions.

We have uploaded a revised version, with red margin notes naming the review point each change addresses.

Q1 (one-step motivation)

The question comes from how a remaining budget is used. Under GDP, an analyst who has spent G1G_1 against a cap G2G_{\sqrt2} passes on the residual G1G_1 alone, and the next Gaussian request is tested against it without the original pair. A residual also describes the whole admissible set before any request is named. We ask when exact profile accounting allows the same hand-off for arbitrary requests: one experiment R⋆R^\star such that C⊗R⪯BC\otimes R\preceq B exactly when R⪯R⋆R\preceq R^\star.

Theorem 1 shows that an exact hand-off at every feasible cap is possible exactly for erasure-type spends, those with a single finite privacy-loss value, possibly with revealing mass. Gaussian, Laplace, subsampled Gaussian and randomized-response mechanisms lie outside this class, and each has a cap at which every safe residual rejects a request that the direct test accepts. Example 1 and Table 2 quantify this for the Gaussian pair above: RR(0.70) is admissible, the residual G1G_1 allows at most RR(0.6915), and no feasible residual admits both G1G_1 and RR(0.70).

W1 (writing and framing)

The spend CC bounds past disclosure and the cap BB bounds total disclosure. "Preserving every feasible choice of the next release" is the equivalence in Q1: R⋆R^\star is the greatest element of the feasible set, which is stronger than being maximal. The revised abstract and introduction define each term at first use, and Section 3 separates greatest from maximal.

On Theorem 3, the revised Section 5 opens by comparing two properties of composition with a fixed spend. Equivalence cancellation recovers R1≃R2R_1\simeq R_2 from C⊗R1≃C⊗R2C\otimes R_1\simeq C\otimes R_2. Order reflection recovers R1⪯R2R_1\preceq R_2 from C⊗R1⪯C⊗R2C\otimes R_1\preceq C\otimes R_2. A remaining budget needs the second property. A finite-support spend with several finite loss values has only the first, and this gap yields a cap with no greatest residual.

W2 (technical contribution), W3 (narrative)

The criterion of Regehr et al. (their Theorem 7, stated as Proposition 2 and attributed in Section 4) reduces existence at every cap to one property of the spend: composition with CC preserves every nonempty supremum. The paper adds three results, listed in the last paragraph of the revised introduction; the first lies outside that criterion.

First, the product-cap obstruction of Theorem 3 does not pass through the supremum criterion (Section 5.1, Appendices D.6 and D.7). The criterion's gap is a strict inequality at one threshold, at a supremum cap, for requests with singular mass. A product cap needs hR1>hR2h_{R_1}>h_{R_2} somewhere and hC⊗R1≤hC⊗R2h_{C\otimes R_1}\le h_{C\otimes R_2} at every threshold. For every spend with at least two finite loss values, at arbitrary locations and weights, the proof builds a sign-changing sum of bumps on the log-threshold axis in which each positive translate created by composition is offset by a negative bump placed for it. It then perturbs a likelihood-ratio density by ±(ρ/2)d′′\pm(\rho/2)d'', which keeps mass and mean one, so the requests are valid, bounded and mutually absolutely continuous. At B=C⊗R2B=C\otimes R_2 a greatest residual would satisfy

C⊗R2⪯C⊗R⋆⪯B=C⊗R2,C\otimes R_2\preceq C\otimes R^\star\preceq B=C\otimes R_2,

and Torgersen cancellation gives R⋆≃R2R^\star\simeq R_2, which the feasible R1⪯̸R2R_1\not\preceq R_2 excludes.

Second, Theorem 1 decides the criterion for every spend (Appendix D.3). Let WW be e−LCe^{-L_C} conditioned on finite loss. For an arbitrary nonconstant law of WW, the proof takes a point aa with mass on both sides and builds two requests A,DA,D with hA>hDh_A>h_D below aa and hA<hDh_A<h_D above it. The crossing lies inside the given law, so

Emax⁡{hA(W),hD(W)}>max⁡{E hA(W),E hD(W)},\mathbb E\max\lbrace h_A(W),h_D(W)\rbrace>\max\lbrace \mathbb E\ h_A(W),\mathbb E\ h_D(W)\rbrace,

and A∨DA\vee D is infeasible at the cap (C⊗A)∨(C⊗D)(C\otimes A)\vee(C\otimes D). The requests are fitted to the law of the spend, which turns the criterion into an equivalence over all spends. Conversely, one finite value makes composition a single rescaling of the threshold, which commutes with suprema, with revealing mass allowed.

Third, Theorem 4 classifies whole additive families (Section 6, Appendix D.8). Finite support at each time, with no uniform bound, forces one finite value: a two-point law at time t/nt/n would give n+1n+1 values at time tt. Subtraction is then exact for every request, also outside the family.

The revised introduction and conclusion follow this narrative: an exact single remaining budget exists exactly for erasure-type spends, Section 5 shows how it fails elsewhere, and Section 6 extends this to additive families.

We hope these responses address your concerns, and we would be grateful if you would consider raising your score.

Thank you for the response. Unfortunately I still cannot understand the motivation for or signficance of the results, so I will maintain my score.

Thank you for reading our response and for letting us know. In brief, the paper asks when a partly spent privacy budget can be passed on as a single remaining budget that decides every later request correctly. Under Gaussian DP this is standard, since the remaining budget is one Gaussian parameter. Under exact accounting, we show it is possible at every cap only for erasure-type spends; for Gaussian, Laplace, subsampled Gaussian and randomized-response spends there are caps at which any remaining budget rejects a request that the full history would admit.

Thank you again for the time and care you have given to our paper. We will take your comments into account in making the motivation and significance clearer, and we would be glad to answer any further questions.

Reviewer_Kwi4OpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies whether the "remaining privacy budget" in differential privacy (after some budget has already been spent under a total cap) can be summarized as a single benchmark mechanism that preserves all valid choices for the next query. Using exact privacy profiles, the authors prove an impossibility result: a single exact remaining-budget representation exists for all caps only when the already-spent mechanism has at most a single finite privacy-loss value (essentially a trivial erasure/blatant-reveal mechanism). For almost all standard mechanisms—including Gaussian DP and Randomized Response—any single remaining-budget representation will inevitably exclude some valid next mechanisms. The authors illustrate this with an exact finite example and a concrete Gaussian DP example where the standard Gaussian remaining budget excludes a valid Randomized Response query.

Critical strengths
  • Easy to follow: The paper is well-organized and walks through the main question step by step. The concrete numerical examples (the rational example in Section 5.2 and the Gaussian vs. Randomized Response example in Section 7) make the main takeaway easy to understand.
  • Interesting impossibility result with some novel pieces: While related to recent work on privacy filters, the paper provides a few clean additions of its own, including a construction showing that composition does not preserve the ordering of mechanisms for multi-value spends (Section 5), a classification of finite-support composition families (Section 6), and a concrete calculation showing that a Gaussian remaining budget rejects a valid Randomized Response mechanism (Section 7).
Critical weaknesses
  • Heavy reliance on Regehr et al. (2026) without clear attribution of what is new: The paper relies heavily on Regehr et al. (2026), citing five different results from that work to set up and prove its main theorem. Regehr et al. already defined the same one-step residual update, showed that exact profile filters fail even in a single step under a Gaussian budget, and noted that single-value privacy loss is the degenerate exception. The submission uses their results as building blocks without clearly explaining what part of the main characterization (Theorem 1) is genuinely new versus a direct restatement of Regehr et al. in the one-step setting.
  • Unclear practical takeaway for both non-adaptive and adaptive settings: In the non-adaptive one-step setting studied here, the lack of a single remaining-budget mechanism is easy to work around: an analyst can simply keep both the spent budget and the total cap and check their composition directly (which standard numerical privacy accountants already do). Even in the paper's own Wine dataset experiment (Section 7.1), direct checking improves accuracy by only 0.02% over using the Gaussian residual, and both private options perform worse than a simple non-private baseline. In the multi-step adaptive setting (where keeping a single residual actually matters), the question was already resolved by Regehr et al. (2026) and Tran et al. (2026), who showed that the "direct admission" check recommended in Section 7 becomes unsafe when applied across multiple adaptive rounds.
Questions for authors
  1. Could you clearly state what Theorem 1 and Proposition 2 add beyond specializing the results of Regehr et al. (2026) (specifically their Theorem 7, Corollary 4, and Lemma 4) to a single fixed spend?
  2. In a single non-adaptive step, why not simply store the spent budget and cap as a pair rather than compressing them into a single residual mechanism?
2 comments

Thank you for your detailed review. We appreciate your helpful suggestions.

We have uploaded a revised version, with red margin notes naming the review point each change addresses.

W1 (attribution), Q1

Proposition 2 is the fixed-spend form of Regehr et al.'s Theorem 7 and is attributed to them in Section 4. The one-step residue update, the supremum construction and the gap mechanism are theirs as well, and the revised introduction ends with a paragraph that states this. Their criterion reduces existence at every cap to one property of the spend: composition with CC preserves every supremum of requests. Their Lemma 4 obtains a gap for at least one of two incomparable nondegenerate PLDs, and their Corollary 4 establishes failure under some Gaussian budget. The additions are in Theorems 1, 3 and 4.

First, Theorem 1 decides the criterion for each fixed spend with an arbitrary privacy-loss law: an exact remaining budget exists at every feasible cap exactly when the privacy loss, whenever finite, takes a single value. For every nonconstant finite-loss law, Appendix D.3 builds two requests whose profiles cross at a point with mass of the law on both sides. Composition averages the profiles over the law, so the spend composed with the supremum of the two requests exceeds the cap formed from the two products. In the other direction, a single value makes composition one rescaling of the threshold, which commutes with suprema.

Second, Section 5 separates equivalence cancellation from order reflection (Theorem 3). The gap in the first step is a strict inequality at one threshold, and a product cap needs an inequality at every threshold. Section 5.1 constructs, for every finite-support spend with several finite loss values, two bounded, mutually absolutely continuous requests whose profiles cross while their products with the spend are ordered at every threshold. Torgersen cancellation then shows that the larger product is a cap with no greatest residual, and Section 5.2 gives an exact rational instance.

Third, Section 6 classifies weakly continuous semigroups with finite privacy-loss support at every time (Theorem 4): they have one finite loss value at each time and are erasure semigroups in the self-dual case. The criterion concerns one spend, and this result concerns a whole additive family.

W2 (practical takeaway), Q2

Passing on a remaining budget is how budgets move between accounting steps. Under GDP, after spend G1G_1 against cap G2G_{\sqrt2} the analyst passes on G1G_1 alone, and the next Gaussian request is compared with it without the original pair (Section 1, first paragraph). A residual also describes the whole admissible set before any request is named, whereas the direct test answers for one specified request (Section 1, second paragraph). The paper determines when this interface is lossless and what it costs otherwise.

For erasure-type spends, one experiment reproduces every admission decision at every cap, and in additive families of this type subtraction is exact for every request (Theorems 1 and 4). For every other spend, Gaussian and randomized-response spends included, there is a cap at which every safe residual rejects a request that the direct test C⊗R⪯BC\otimes R\preceq B accepts, so the pair (C,B)(C,B) is the record that loses no admissible request (Table 1 and the paragraph after it).

Example 1 quantifies this for subtraction under GDP. With spend G1G_1 and cap G2G_{\sqrt2} the direct test admits RR(0.70), the residual G1G_1 allows at most RR(0.6915), and no feasible residual admits both G1G_1 and RR(0.70) (Table 2). The bit error is 30.000% under the direct test and 30.854% under the Gaussian residual (Table 3), and the residual channel equals RR(0.70) followed by an extra output flip with probability 2.134% (Section 7.1).

On adaptive rounds, C⊗RC\otimes R denotes independent composition, and transcript-dependent rounds are treated by Regehr et al. and Tran et al. (Section 2). The one-step classification shows that the loss of a subtraction-style budget is a property of the spend, present for a single independent release.

The Wine illustration in Section 7.1 shows how the channel difference passes to a task. By Eq. (3) the task difference is the channel improvement times the selector's signed utility margin, so admission fixes the coefficient and the task fixes the margin. There it is 0.0215 points, reported together with the public fixed-option rule; the non-existence result rests on Example 1.

We hope these responses address your concerns, and we would be grateful if you would consider raising your score.

Thanks for the response and for updating the paper. Clarifying what comes from Regehr et al. versus your own results in Theorems 1, 3, and 4 is helpful. That said, I am still unsure of the usefulness of this result so I will keep my original rating.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.