acceptodds
Under review as a conference paper at ICLR 2027

Support Selection Beyond Smooth DAG Exactness: Completion Geometry, Score Margins, and Selective Certificates

Abstract

Continuous DAG learning often relies on a smooth exact constraint to enforce acyclicity. Exactness specifies the zero set, but leaves two questions unanswered. How informative is a small nonzero residual? Which edges should be removed when several repairs restore acyclicity? For the first question, we prove sharp local limits for sufficiently smooth exact representations on signed candidate neighborhoods of a fixed DAG. The bounds reveal that the smallest minimal cycle completion controls the earliest possible response, while the largest controls the best uniform distance-error exponent. They also distinguish signed or vector residuals from nonnegative scalars. For the second question, we show how a score margin breaks a feasibility tie in an isolated-cycle flow. We then use simultaneous parent-set score intervals to account for sampling uncertainty and certify features shared by all population optima of a frozen score, without requiring a unique graph. Controlled experiments recover the predicted orders and verify certification under nonunique optima, while comparisons with standard optimizers show the flow model’s limitations. The results clarify what changing a smooth constraint can improve and what requires evidence from scores and data.

Primary area:
causal reasoning
Keywords:
causal discovery, DAG structure learning, continuous optimization, acyclicity constraints, support selection, selective inference

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_P8p2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3(+2)
Soundness
2
Read review
Paper summary

This paper deals with two interesting questions within the continuous DAG learning domain: How informative is a small nonzero residual? Which edges should be removed when several repairs restore acyclicity? For the first question, this authors show sharp local limits for sufficiently smooth exact representations. For the second question, this authors show how a score margin breaks a feasibility tie in an isolated-cycle flow. The experiments verify the proposed ideas.

Critical strengths
  • The theory part of this paper is rather strong: Theorem 2.1 gives a clean, formula-independent restriction on the Taylor structure of any smooth exact residual; Theorem 2.2 formalizes the known gradient degeneracy of NOTEARS and DAGMA at the origin. Theorem 4.1 gives a finite-sample guarantee that does not need a unique optimum.
  • The task itself is important.
  • The paper is honest and frank, even reporting some unfavorable findings: the flow ordering does not transfer to Adam or L-BFGS-B, the initial ratio explains the predictor gain in Table 1, and 23 of 80 certified absent adjacencies disagree with the true graph.
  • The exact-population calibration is a well-designed sanity check of the certificate, and the Lean formulation provides more trust.
Critical weaknesses
  • This paper is presented very badly: 1) there are many theorems, but not providing some brief sketch or intuitive examples for the readers to quickly capture the main ideas. 2) Table 1 and Table 2 are sparsely presented, leaving a lot of empty spaces, maybe using wraptable can be helpful. 3) Figure 2 is also not so informative, only 4 points in one line, giving shallow reliability.
  • The theory part needs more refinement, e.g., Theorem 2.1 is essentially a corollary of vanishing on coordinate subspaces, and when candidates include all missing edges, q=1q=1, so the bounds say only O(t2)O(t^2) and carry no information.
  • Please give a concrete case where the q or r analysis changes the choice of constraint or stopping rule?
Questions for authors

See above.

4 comments
AuthorsOpenReview ↗

Response to Reviewer P8p2

Thank you for reviewing our paper.

1. What does Theorem 2.1 add beyond vanishing on coordinate subspaces?

We apologize for not clearly explaining this part in the text. Coordinate vanishing does indeed directly provide a qq-order bound for signed or vector residuals. However, the 2q2q-order bound for nonnegative scalars requires another step. If there exists a first nonnegative, non-zero homogeneous Taylor term in the signed neighborhood, then each Newton vertex monomial should involve at least qq candidate edge weights, and all exponents must be even; otherwise, we only need to change the sign of one edge weight to make the dominant monomial negative. Therefore, the total degree is at least 2q2q. Furthermore, it is significant that this argument does not require the assumption (W∘WW \circ W), and the constructions in the theorem in the text also achieve both bounds.

2. What remains informative when all missing edges are allowed and q=1q=1?

q=1q=1 indicates that for the easiest cycle to form, we only need to add one edge. However, there may be other ways to form cycles in the same graph. We added a four-node example on page 4 of the revised version. If the original graph has only one edge 1→21\to2 with a weight of 11, and we take a very small value t>0t>0 and only add weight tt to the reverse edge 2→12\to1, then the EXP residual is approximately t2t^2. However, if we add weight tt to 2→32\to3, 3→43\to4, and 4→14\to1, the residual is only about t6/6t^6/6. In both cases, we need to delete an edge with a weight of tt to break the cycle. That is, we actually need to change the same amount, but the residual can be much smaller. In the second case, the residual hh is approximately t6/6t^6/6, so the repair distance tt is approximately (6h)1/6(6h)^{1/6}. If a larger exponent is used to estimate the distance, when tt is sufficiently small, the calculated upper bound will actually be smaller than the actual change needed. This is why, in this example, although q=1q=1, the cycle completed by the three edges r=3r=3 still restricts the exponent of the uniform distance upper bound to 1/61/6.

3. How would these results change a penalty choice or stopping check?

If we follow the path that reaches the sharp order, the response will be pushed from order 2q2q to order 4q4q due to the square of the nonnegative scalar residual. Table 11 in the paper provides a direct comparison of this. When we set q=6q=6 and perform 1,0001,000 projected gradient steps, and then use the vector residual norm as a penalty, the median of the remaining edge weights is 0.1220.122. However, if we use its square, the result is 0.4060.406. Therefore, for the stopping decision, we can first determine the maximum edge weight δ\delta that we are willing to delete, and then check whether the remaining edges form a DAG. That is, τ(W)≤δ\tau(W)\le\delta in equation 141141. However, if we want to calculate the residual tolerance from δ\delta, then in addition to rr, we also need the error bound constant and the effective neighborhood to achieve this goal.

4. Typography and presentation.

We have carefully considered your feedback and revised the layout and representation of the tables and figures. In Figure 2, the four points on each curve actually correspond to four different sample sizes. Each point is based on 200200 independent replicate experiments, and we have also provided point-by-point, two-sided 9595% exact binomial intervals. We have also optimized the writing style, adding some proof sketches and simple examples to improve the readability of the paper. These revisions will be presented in the revised version.

We sincerely appreciate your time and effort in reviewing our manuscript and offering valuable suggestions. As the author-reviewer discussion phase is drawing to a close, we would like to confirm whether our responses have effectively addressed your concerns. If you require further clarification or have any additional concerns, please do not hesitate to contact us. We are more than willing to continue our communication with you.

Thanks for the detailed response and the revision. The added proof sketches, the four-node example, the confidence intervals in Figure 2 and the improved layout make the paper much easier to follow. My concerns regarding the presentation have been addressed, and I will raise my Presentation score accordingly.

However, I think the theoretical part still needs further refinement. Two of my questions are not directly answered:

  1. The case q = 1. When all missing edges are candidates, q = 1, and Theorem 2.1 only gives O(t²). The response moves to r instead. What does Theorem 2.1 itself tell us in this setting? Also, r can be as large as d here, so the bound of Theorem 2.2 becomes 1/(2d). Is this bound still informative?
  2. Practical use of q and r. The stopping check τ(W) ≤ δ does not use q or r, and the constant K cannot be computed in practice. Could you give one concrete case where knowing q or r changes what a practitioner does, for example which constraint to use or when to stop?

Therefore, I keep my overall rating for now.

AuthorsOpenReview ↗

Response to Reviewer P8p2

Thank you for reviewing the revision and for recognizing the improvements in presentation. Our previous reply moved too quickly from qq to rr. We appreciate the opportunity to clarify what each result tells us.

1. What does Theorem 2.1 itself tell us when all missing edges are candidates and q=1q=1?

We agree that O(t2)O(t^2) alone says little. Theorem 2.1 also tells us which candidate edges must appear together for the residual to respond. Lemma 2 of [1] gives a related return-path criterion for trace constraints in nonnegative adjacency coordinates. In our example on page 4, the base graph has a single unit-weight edge 1→21\to2. Only 2→12\to1 can close a cycle on its own. By our Proposition D.10, the quadratic term of any smooth nonnegative exact residual can therefore only be c x2→12c\,x_{2\to1}^2, with c≥0c\geq0.

Now give edges 2→32\to3, 3→43\to4, and 4→14\to1 weight tt each and set the other candidate weights to zero. All three edges are needed to close the cycle. The candidate set is the same in both cases. The reverse edge alone accounts for q=1q=1. Along the three-edge perturbation, Theorem 2.1 gives O(t6)O(t^6), so no term of degree below six can appear. EXP attains this order, with leading term t6/6t^6/6.

On the question of dimension, when all missing edges are candidates and the base graph has at least one edge, r≤d−1r\leq d-1. A base graph with a single edge attains this bound, so the optimal EXP distance exponent is 1/[2(d−1)]1/[2(d-1)] and the guarantee does weaken as dd grows. The case r=dr=d occurs when the base graph has no edges, in which case q=2q=2.

To see which graphs have a stronger guarantee, let ww be the largest number of vertices in the base graph such that none can reach another. By the strict alternating-cycle property in [2], the tails of the edges in any minimal completion of size at least two are pairwise unreachable. Hence r≤wr\leq w. With all missing edges available, joining those ww vertices in a directed cycle attains the bound. If w=1w=1, a single back edge suffices. Thus r=wr=w in our setting. For example, two disjoint chains have w=2w=2. By Theorem 2.2, their optimal EXP distance exponent is 1/41/4, however long the chains are.

2. How does knowing rr change the stopping decision?

Appendix C of [3] measures the smallest edge-weight threshold that makes a given matrix acyclic. Here we fix the threshold first. Suppose we are willing to remove edges of magnitude at most δ=0.05\delta=0.05. Is 10−810^{-8}, the default residual tolerance in the linear NOTEARS implementation [4], sufficient? Consider two four-node graphs. In both, the base edges have unit weight and all missing edges are candidates.

The first base graph contains all six forward edges in node order, so r=1r=1. Any back edge of magnitude above δ\delta forms a two-cycle with a unit-weight edge. That cycle alone contributes more than δ2=0.0025\delta^2=0.0025 to EXP. A residual tolerance of 10−810^{-8} therefore guarantees acyclicity after thresholding.

The second base graph contains only 1→21\to2, so r=3r=3. If a cycle survives thresholding, there is one of length at most four using at most three candidate edges. In the EXP series, the product of squared edge weights for a four-cycle has coefficient 4/4!=1/64/4!=1/6. Shorter cycles have larger coefficients. A sufficient tolerance is therefore δ6/6≈2.60×10−9\delta^6/6\approx2.60\times10^{-9}.

To see why 10−810^{-8} is too loose for the second graph, give 2→32\to3, 3→43\to4, and 4→14\to1 weight 0.060.06 each and set the other candidate weights to zero. For this four-cycle, the leading contribution to EXP is 4(0.06)6/4!=7.776×10−94(0.06)^6/4!=7.776\times10^{-9}. Including all higher-order terms gives hexp⁡(W)≈7.78×10−9h_{\exp}(W)\approx7.78\times10^{-9}, below the 10−810^{-8} tolerance, while all four edges survive the 0.050.05 cutoff. Both tolerance calculations follow directly from the fixed base graphs and the EXP formula, without estimating KK.

References and code

[1] DAGs with No Fears: A Closer Look at Continuous Optimization for Learning Bayesian Networks. NeurIPS 2020, Lemma 2, p. 3.

[2] Dimension, Graph and Hypergraph Coloring. Order, 2000, pp. 170–171.

[3] DAGs with NO TEARS: Continuous Optimization for Structure Learning. NeurIPS 2018. See Appendix C and Fig. 5 in the linked arXiv version, pp. 17–18.

[4] Official linear NOTEARS implementation. Default h_tol in notears_linear.

Reviewer_CNkaOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2(+1)
Soundness
3
Read review
Paper summary

The paper asks what the smooth acyclicity constraints used in continuous DAG learning, such as NOTEARS and DAGMA, can and cannot tell us.

First, it shows that a small constraint value does not mean the learned graph is close to a DAG. This limit comes from the structure of the graph, not from the choice of constraint: cycles that need several weak edges to close are nearly invisible to any smooth constraint, so designing a better formula cannot fix it.

Second, it argues that when several edge removals would break a cycle, the constraint cannot decide between them, so the data score must make that choice. It illustrates this with a simplified gradient flow on a single cycle.

Third, because the score is estimated from data, its preference may be noise. The paper therefore proposes a procedure that certifies only those graph features, such as an adjacency or a collider, that are shared by every graph the data cannot rule out, without needing a single best graph.

Small controlled experiments check the theory, and the certificate is applied to graphs proposed by several continuous learning methods.

Critical strengths
  1. Part 1 turns several scattered observations about smooth acyclicity constraints into one clear insight. How well a constraint can detect cycles, and how well a small constraint value guarantees that the graph is close to a DAG, is limited by the structure of the graph, not by the choice of formula. The paper also shows these limits are tight. The takeaway for the field is useful: designing new constraints cannot remove the dependence on graph structure.
  2. The claims are carefully scoped, and parts of the proofs are formally verified.
Critical weaknesses
  1. Limited significance. I found no clear errors. However, the results concern simplified settings: the neighborhood of a fixed DAG, a gradient flow on a single cycle, and the best graphs under a modified score. They do not yet give practitioners something to use. For example, Part 1 does not lead to a usable stopping rule, which the authors acknowledge.

  2. Parts 2 and 3 add little and are loosely connected to Part 1.

    • The ordering predicted by the flow analysis does not carry over to standard optimizers such as Adam and L-BFGS-B, as the authors’ own experiments show.

    • In Part 3, continuous methods only propose candidates. The main certificate is a discrete search over parent sets. It does not use the smooth constraint, and Part 1’s bounds play no role in it.

    • The idea of certifying features shared by all near-optimal graphs is close to prior work that is not discussed: enumerating near-optimal networks (Liao et al., AAAI 2019) and bootstrap confidence for network features (Friedman, Goldszmidt & Wyner, UAI 1999).The finite-sample guarantee appears to be new, but a comparison with these approaches would clarify what the certificate adds.

  3. Clarity. The paper is hard to follow. The abstract and introduction rely on terms explained only later, and the main messages are never stated plainly. The technical links between the three parts are thin.

Questions for authors
  1. What is the single main contribution, and why do Parts 2 and 3 belong in the same paper as Part 1?
  2. What should practitioners do differently because of Part 1?
4 comments
AuthorsOpenReview ↗

Response to Reviewer CNka

Thank you for reviewing our paper.

1. What is the main contribution, and why do the three parts belong together?

In Section 2 of this paper, we explored a fundamental question: if the acyclicity residual is already small, how much modification is needed to completely break the cycle? Our core contribution is proving which edges can be combined to form a cycle and defining an upper limit on this distance difference. More importantly, this constraint holds universally for any smooth and non-negative acyclic constraint.

The following sections naturally follow up with the question: given multiple possible edges, which one should be deleted? For example, in our four-node example, deleting 3→43\to4 or 4→14\to1 breaks the cycle, and both residuals become zero. Section 3 discusses under what circumstances data scoring will cause one edge to shrink faster than the other. However, our experiments also show that the slower trajectories predicted by the gradient flow may not actually run slower with Adam or L-BFGS-B optimization, highlighting the inherent gap between idealized continuous dynamics and discrete optimizer heuristics.

Finally, Section 4 answers another crucial question: even if the algorithm ultimately leaves an edge, is the existing data sufficient to support our claim and reporting of it? You are absolutely right; this step is accomplished by performing independent holdout comparisons between candidate graphs. Logically, it doesn't rely on the bounds derived in Section 2, but it is indispensable in the edge selection process. Our narrative proceeds from theoretical limits to practical optimization and, ultimately, to statistical verification.

2. What should practitioners do differently after Section 2?

In practical applications, we recommend reporting the minimum edge-removal threshold τ(W)\tau(W)—sufficient to eliminate all directed cycles—alongside the residual. Relying solely on a fixed residual tolerance (e.g., 10−1210^{-12}) is often misleading, as cycles of different lengths exhibit drastically different sensitivities to residuals. For example, as shown in Table 10, under the same 10−1210^{-12} residual threshold, the edge weights in a 22-cycle decay to 0.00100.0010 (well below the 0.050.05 tolerance limit), whereas those in an 88-cycle remain as high as 0.30420.3042.

Therefore, we advocate a more robust workflow: first defining the maximum tolerable perturbation magnitude for edge weights—say, setting it to 0.050.05. By directly comparing the threshold τ(W)\tau(W) with 0.050.05, practitioners can rigorously determine whether the permitted edge adjustments are sufficient to break the cycles, bypassing the need to blindly guess or align raw residual values.

3. How does the certificate differ from the work of Liao et al. and Friedman et al.?

Liao et al. retained a pool of candidate graphs that were 'close enough' to the best graph by setting a scoring gap; Friedman et al. assessed the selection frequency of structural features via resampling. Unlike them, we explore a more challenging problem: when multiple graphs have the same overall optimal score, can the existing data support reporting a common feature among them? To address this, after model training, we completely fixed the scoring criteria and the candidate graph family, and used independent holdout data to safely exclude graphs with truly poor scores. As proven by Theorem 4.1, under the given conditions, with probability at least 1−α1-\alpha, all the overall optimal graphs are retained within the confidence family, ensuring that every structural feature we report necessarily holds true on these graphs. We have included a comparison of your two proposed works in the relevant sections and tables of the revised manuscript.

4. Presentation

We have carefully considered your feedback. To improve the readability and coherence of the three parts of the paper, we have optimized the writing style, added proof sketches and simple examples. These revisions will be presented in the revised version.

We sincerely appreciate your time and effort in reviewing our manuscript and offering valuable suggestions. As the author-reviewer discussion phase is drawing to a close, we would like to confirm whether our responses have effectively addressed your concerns. If you require further clarification or have any additional concerns, please do not hesitate to contact us. We are more than willing to continue our communication with you.

Thank you for the clarifications and the revisions. Naming Section 2 as the main contribution, the practical recommendation on edge-removal thresholds, and the added comparison with prior work are all helpful. I have raised my presentation score. However, the rebuttal confirms that Sections 3 and 4 are connected to Section 2 by narrative rather than technical dependence, which remains my main concern. I keep my overall rating.

AuthorsOpenReview ↗

Response to Reviewer CNka

Thank you for reviewing the revision and for raising your presentation score. You are right that the coverage guarantee in Theorem 4.1 does not depend on the order bounds in Section 2. Our previous response left two specific technical connections unclear.

The first connects Sections 2 and 3. On an isolated directed cycle, let pp be the product of the squared edge weights. The expansion in Section 2 gives h=ϕ(p)h=\phi(p) with ϕ′(0)>0\phi'(0)>0 for both EXP and DAGMA. This is the starting point of Proposition 3.1, which uses this form to calculate how long it takes the ratio of two initially nearly equal edge weights to reach a specified value. Under gradient flow driven only by the acyclicity penalty, replacing hh with h2/2h^2/2 changes that time from Θ(ε−1)\Theta(\varepsilon^{-1}) to Θ(ε−3)\Theta(\varepsilon^{-3}), where the two weights start at a+εa+\varepsilon and a−εa-\varepsilon (p. 5).

The second connects Sections 2 and 4. Section 2 identifies minimal sets of edges that complete a cycle. Those same sets determine which deletions are feasible repairs. On a fixed support, deleting a set of edges leaves a DAG if and only if it removes at least one edge from every minimal completion (Lemma G.1, pp. 52–53). When deletion costs are additive, we compare the total costs of these repairs on independent holdout data. Theorem G.2 then tells us which repairs we can rule out as population optima. In this sense, the completion structure determines which repairs enter the comparison, and the holdout data determine which can be ruled out. The more general Theorem 4.1 compares parent-set scores across candidate DAGs. Its coverage guarantee comes from simultaneous bounds on those score differences.

Reviewer_pPN9OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

The paper investigates limitations of smooth acyclicity constraints in DAG learning. Section 2 establishes bounds on how sensitive these constraints can be to small cycles, and on how well the constraint value is able to reflect the distance to an acyclic graph. Section 3 investigates how a data-fitting score can help an optimizer decide which edges to remove when breaking a cycle. Section 4 presents a statistical method for deciding which graph features can be reported with confidence.

Critical strengths

Dissecting the role of the acyclicity constraint in continuous DAG learning is a worthwhile subject.

Section 2 shows that there are fundamental limits to how well smooth acyclicity constraints can detect small cycles and that these limits apply to any smooth exact constraint

The certification idea presented in section 4 seems sensible.

Critical weaknesses

The paper is dense and introduces a large amount of technical material. The main ideas are difficult to follow, and it is not clear how the contributions fit together. It would be beneficial if the paper provided more intuition and examples.

Relatedly, the paper does not consistently present much of a connection between its contributions. In particular, Section 4 does not appear to make direct use of the results from earlier sections.

Section 3 studies a simplified optimizer whose behavior, by the paper’s own tests, doesn’t match real optimizers, so the practical relevancy of section 3 is a bit unclear.

The experiments do not use real-world data.

Questions for authors

What are the main practical implications of Section 3? Given the results with standard optimizers, in which situations would the gradient-flow analysis be useful?

2 comments
AuthorsOpenReview ↗

Response to Reviewer pPN9

Thank you for reviewing our paper.

1. What are the main practical implications of Section 3?

Assume that removing either of the two candidate edges would break the cycle. In reality, the acyclic penalty can indeed drive edge weight reduction, but even if it has almost no bias towards either edge, the data scoring might still cause one edge to shrink before the other. When we actually encounter two edge weights that are quite close, we can calculate the effects of these two factors separately on the same pair of edge weights to see which factor is driving them apart. If the combined effect consistently favors one edge, and the degree of bias is always above a positive constant, then Proposition 3.2 in the paper can provide its upper bound on the time, for example, how long it will take for this edge to shrink to one-tenth the size of the other.

2. Given the results with standard optimizers, in which situations would the gradient-flow analysis be useful?

Gradient flow provides an idealized analytical benchmark when we want to explain why one edge shrinks faster than another. To test its effectiveness in practice, we examined the performance of Adam and L-BFGS-B. The results show that the order of gradient flow predictions is highly sensitive to the choice of solver and the initial scale (e.g., Adam reverses at the 0.40.4 scale, and L-BFGS-B has no stable ordering). This indicates that the theoretical bias of the objective function cannot be directly translated into deterministic behavior of the solver, and relying solely on the path during the optimization process to determine structural evolution is unreliable. Further research into the stability of real solvers will be our future work.

3. Why do these three sections belong to the same paper?

In Section 2 of this paper, we explored a fundamental question: if the acyclicity residual is already small, how much modification is needed to completely break the cycle? The following sections naturally follow up with the question: given multiple possible edges, which one should be deleted? For example, in our four-node example, deleting 3→43\to4 or 4→14\to1 breaks the cycle, and both residuals become zero. Section 3 discusses under what circumstances data scoring will cause one edge to shrink faster than the other. However, our experiments also show that the slower trajectories predicted by the gradient flow may not actually run slower with Adam or L-BFGS-B optimization, highlighting the inherent gap between idealized continuous dynamics and discrete optimizer heuristics. Finally, Section 4 answers another crucial question: even if the algorithm ultimately leaves an edge, is the existing data sufficient to support our claim and reporting of it? Our narrative proceeds from theoretical limits to practical optimization and, ultimately, to statistical verification.

4. Presentation

We have carefully considered your feedback on the writing and formatting. To improve readability, we have optimized the writing style, added proof sketches and simple examples. These revisions will be presented in the revised version.

5. Real-world data experiments

We added experiments using the real-world UCI Abalone dataset in Appendix H.11 of the revised manuscript. We used 1,0001,000 samples from the total sample to fit the scores, and the remaining 3,1773,177 samples for comparison. We selected four variables. There were a total of 543543 candidate DAGs. In the main split, 207207 graphs without the shell length-shell diameter adjacency were worse than the reference graph. Even after accounting for sampling uncertainty, the minimum conservative lower bound of the score difference was still positive (0.005060.00506). Therefore, all 207207 graphs were excluded. Of the 110110 retained graphs, 7474 pointed in one direction and 3636 pointed in the opposite direction, so the program abstained on the direction. Similarly, for "shell height-dry shell weight," although 9696 retained graphs contained this edge, 1414 graphs without this edge could not be excluded by the data, and the program also chose to abstain. In the ten pre-fixed splits of the experiment, the same adjacency was successfully and stably certified each time, and both directions were retained each time. This real-world example directly demonstrates how the method in Section 4 can be tested on real data and stops at what the evidence supports, i.e., reporting only the certified common features when the whole graph is not unique. We will submit the revised paper and the code.

We sincerely appreciate your time and effort in reviewing our manuscript and offering valuable suggestions. As the author-reviewer discussion phase is drawing to a close, we would like to confirm whether our responses have effectively addressed your concerns. If you require further clarification or have any additional concerns, please do not hesitate to contact us. We are more than willing to continue our communication with you.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary of Revisions

We are very grateful to the three reviewers for their careful reviews and constructive suggestions. We have uploaded the revised manuscript and the code package for the newly added experiment. All additions and revisions in the manuscript are highlighted in blue.

We have expanded the proof outlines for Theorems 2.1, 2.2, and 4.1 (pages 4 and 8). On page 4, we have replaced the original algebraic example with a four-node example in which all missing edges are allowed to be added. The two constructions require the same magnitude of edge-weight changes to break the cycles, but their EXP residuals approach zero at different rates. On page 10, we use the results from Table 10 of the original manuscript to illustrate how to determine whether a given edge-deletion threshold is sufficient to break all cycles.

On page 6, we have added the median number of Adam steps needed to reach the target edge-weight ratio at two initial edge-weight scales. These counts illustrate the reversal in ordering between the q=2q=2 and q=4q=4 cases reported in the original manuscript. We have also explained how to compute the separate contributions of the data score and the acyclicity penalty to the relative shrinkage of two edges along a single-cycle gradient flow.

On pages 7–8, we further clarify that the guarantee in Section 4 follows from simultaneous score comparisons using independent holdout data. We have also added comparisons with the work of Liao et al. and Friedman et al. (pages 3 and 38), transposed Table 1 (page 6), and added 9595% exact binomial confidence intervals for each point in Figure 2 based on the original experimental results (page 9).

We have added a real-data experiment using four measurements from the UCI Abalone dataset. In the main data split, all 110110 retained DAGs contain an edge between shell length and shell diameter. Among them, 7474 orient the edge from shell length to shell diameter, while the remaining 3636 orient it in the opposite direction. Thus, under the chosen scoring criterion, the adjacency can be certified, but its direction remains unresolved. This illustrates that, in practice, the method abstains from orienting an edge when the available evidence is insufficient (pages 10 and 71, Appendix H.11).

AuthorsOpenReview ↗

Summary of Revisions

Thank for the continued discussion. We have uploaded a second revision incorporating this discussion. On page 4, we clarify what Theorem 2.1 tells us when q=1q=1. On page 10, we give two EXP residual tolerances for different values of rr under the same edge-removal threshold. On pages 5 and 7, we explain the technical connections between Section 2 and Sections 3 and 4. Appendix I (pages 72–73) characterizes rr when all missing edges are candidates and provides the derivations of these tolerances. All additions and revisions in this update are highlighted in blue, and the highlights from the previous revision are retained.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.