acceptodds
Under review as a conference paper at ICLR 2027

Count-Conserving Diffusion for Point Processes in Time and Space

17 views

Abstract

Planning often needs synthetic event scenarios that meet exact counts in several regions. Existing generative models control such counts at best on average, and correcting their samples afterwards either leaves overlapping counts wrong or meets them by rewriting the sample. We formalize this problem and develop a theory of count conservation for point-process diffusion. Viewing branching diffusion as a Wasserstein–Fisher–Rao gradient flow, we derive the projection of this flow that keeps every expected count, and show through the quadratic variation of the counts that its births and deaths still let each sample's counts drift. Conservation on every sample path requires every jump to lie in the null space of the count map. On a table of cell counts, the target factorizes into a table law and a position law shared by all budgets, and this condition admits three kinds of count-neutral updates and a Markov-basis chain that provably converges under overlap. Count-Conserving Diffusion is built from these updates alone: every sample meets all counts at every step, and one network serves every budget. On California wildfire scenarios with 33 overlapping coverage areas, global projection of DSTPP samples rewrites a median of 68–84% of the output events. Our samples meet every count without correction and are closer to the real data distribution.

Primary area:
learning on time series and dynamical systems
Keywords:
spatio-temporal point processes, diffusion models, constrained generation, exact count constraints, pathwise conservation, conditional generation, Markov bases, Wasserstein--Fisher--Rao

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

This work addresses the generation of spatiotemporal event configurations under predefined count constraints over multiple regions that may overlap. The authors differentiate between the conservation of expected event counts and path-wise conservation for individual samples. The target conditional distribution is factorized into a distribution defined on feasible count tables, paired with a positional distribution conditioned on each such table. The proposed approach integrates a fitted Poisson-mixture table model with count-preserving position diffusion. For scenarios with overlapping constraints, the sampling procedure interleaves position denoising steps with null-space Metropolis–Hastings (MH) updates, which incorporate Markov-basis moves. Experiments on synthetic datasets, earthquake records, emergency call data and wildfire ignition events demonstrate performance gains compared with pipelines that first generate samples and then apply repair or projection post-processing.

Critical strengths

Joint budgets lock the full fine-grained count table, so performance gains under this setup serve as evidence validating the efficacy of the within-cell position model. The wildfire experiment further contrasts direct constrained sampling against post-generation projection with an identical network architecture, which corroborates the practical merit of the sampling trajectory itself. These control experiments reinforce the empirical arguments of the paper. Relevant content can be found on pages 8–9, Sections 5.2–5.3 and Figures 3–4.

Critical weaknesses

Proposition 4 proves that wref(x∣b)w_{\mathrm{ref}}(x\mid b) remains invariant under table updates with fixed proposal weights. Nevertheless, Algorithm 1 keeps updating the table throughout reverse denoising and uniformly resamples relocated events. At intermediate noise levels, the positions conditioned on a given table are typically non-uniform. Consequently, valid Metropolis–Hastings updates for the table do not inherently preserve the target joint distribution over tables and positions.

Even if the reference table probabilities are perfectly accurate and the reverse position kernels are exact, Proposition 4 alone cannot guarantee the correctness of this interleaved sampling scheme. The authors note on page 21 that the final position distribution may depend on the complete trajectory of table states. However, Equation (13) decomposes the total error yet fails to account for the extra positional bias introduced by interleaving table and position updates. Given that overlapping constraints constitute a core use case of this work, this theoretical gap substantially impacts the overall evaluation. Relevant discussion can be found on page 7, Section 4.3; pages 20–21, Appendices D.3–D.4; and Algorithm 1, line 9 on page 24.

Questions for authors

What joint target distribution corresponds to the interleaved table updates and uniform positional resampling steps described in Algorithm 1? If these update operations fail to exactly preserve the intermediate joint distribution, what theoretical analysis or empirical evidence demonstrates that the introduced extra bias is adequately small? Under the same computational budget allocated to table updates, do the main conclusions still hold if we first fully run table sampling to convergence, then fix the obtained table and reinitialize positions before performing position denoising?

2 comments
AuthorsOpenReview ↗

Response to Reviewer QQTb

Thank you for your thorough review and constructive feedback. We address your concerns below and hope these clarifications will help you re-evaluate and update the score:

We adopted your table-first sampler as the default, and at an equal proposal budget its results match the interleaved schedule (Table S3, Appendix E.3).

1. Joint target and interleaving

"What joint target distribution corresponds to the interleaved table updates and uniform positional resampling steps described in Algorithm 1?"

You are right that interleaving does not generally preserve the intended joint law of tables and noised positions, since Proposition 4 establishes invariance of the table distribution only. At intermediate noise levels positions given a table are not uniform, so uniformly relocating events after a table update need not keep that joint law.

Eq. (13) remains an exact decomposition of the actual output distribution, and its position term includes any interleaving error but does not isolate it. We adopt your table-first sampler, which makes the target explicit.

2. Revised sampler

"first fully run table sampling to convergence, then fix the obtained table and reinitialize positions before performing position denoising"

We have made this table-first variant the default and revised Algorithm 1 accordingly:

  1. Start from a feasible table with uniform positions. At k=Kk=K, freeze the table-proposal weights from the first network prediction.
  2. Run the table chain with the same total number of proposals as the original sampler.
  3. Redraw all positions from QxQ_x, uniform positions within the cells of the resulting table xx.
  4. Denoise from k=Kk=K to 11, keeping xx fixed.

The independent redraw removes the position kernel's dependence on the table trajectory, so the output law has the form of Eq. (3),

PTFb=∑xwLb(x)Rθ(⋅∣x),P_{\mathrm{TF}}^b = \sum_x w_L^b(x) R_\theta(\cdot\mid x),

with wLbw_L^b the table law after LL proposals and RθR_\theta the output law of the fixed-table position sampler. By Proposition 3, the error in Eq. (13) splits exactly into a table term and a position term. A chain that is too short can therefore only affect which table is chosen, not where events are placed within it.

3. Matched-budget comparison

"Under the same computational budget allocated to table updates, do the main conclusions still hold"

Yes, at equal table-proposal counts the two samplers give nearly identical results (Table S3, Appendix E.3). We compare them using the same five trained networks, 156 test weeks and five samples per week per network. Entries are interleaved → table-first, and "both" covers air-attack and helitack areas.

MetricWeekly, air-attackWeekly, bothDaily, both
Distance0.883 → 0.8800.857 → 0.8590.465 → 0.468
MMD0.076 → 0.0750.066 → 0.0670.090 → 0.088
Disk-query error2.99 → 2.972.53 → 2.552.35 → 2.33
Intersection-query error2.98 → 3.012.68 → 2.692.41 → 2.43
Busiest-day error1.59 → 1.581.63 → 1.630.84 → 0.85

For weekly counts over both base types, the paired distance difference is 0.002, with a 95% paired block-bootstrap interval of [−0.004, 0.007]. The main comparisons in Table 3 remain unchanged. For example, table-first distance is 0.859 versus 0.980 for DSTPP + global projection.

4. Does the table chain mix adequately?

In our checks, the chain targets the right table law, and a longer chain leaves the results unchanged. To check whether the table chain mixes adequately within the finite proposal budget, we looked at three things.

  • Exact convergence on a small budget. Figure 6 (Appendix D.4) computes the exact total-variation distance on a three-area budget. There, paired moves alone stall, while mixing in basis moves with probability 0.1 gives geometric convergence.
  • Implementation check. Exact enumeration on small feasible sets confirms that the implemented chain keeps the target invariant (Appendix D.4).
  • Longer chains on wildfire. The reported metrics do not change at four times the proposal budget. Table-first MMD is then 0.074, 0.068 and 0.089 in the three settings, within run-to-run variation.

Formally, with fixed proposal weights, Proposition 4 shows that the chain keeps our fitted table law wref(⋅∣b)w_{\mathrm{ref}}(\cdot\mid b) invariant. With Markov-basis moves, it also converges geometrically from every feasible start.

Partition-budget sampling already fixes the table before position denoising, so Tables 1–2 are unaffected. We have updated Algorithm 1, Appendix D.3, and Table 3 accordingly.

Dear Reviewer QQTb,

Thank you again for your careful reading of the sampler. Before the discussion period closes, we wanted to briefly check in on our revision.

Following your suggestion, we made the table-first sampler the default in Algorithm 1. Positions are now redrawn after the table chain finishes, so the joint target is explicit and the error in Eq. (13) splits into a table term and a position term.

At the same table-proposal budget, it gives nearly the same results as the interleaved sampler (Table S3). Our response above gives the revised algorithm and the full comparison.

If you have a moment, we would value your view on whether this resolves the theoretical gap you raised, and we are happy to discuss any remaining point.

Best regards,

The Authors

Reviewer_5WA7OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents the problem of count constrained generation for STPP models on potentially overlapping regions. By introducing a grid for branching diffusion, the method simplifies the task of count per region modelling to a simpler linear system over non-negative integers. The authors show that projecting branching diffusion keeps point counts correct only on average, because every birth and death adds randomness. Thus by separating the count from position, the method first picks how many points go in a cell, and then diffuses the point positions. They show results on multiple datasets and configurations.

Critical strengths
  • Rigorous mathematical derivation and problem definition.
  • Good experiments.
  • Honest treatment of the method’s own limitations (paired-move connectivity), with ablations to show the contribution from the architecture.
Critical weaknesses

Impact: The motivating application is very narrow. On its own this would not be an issue, but combined with the scope of the method it substantially limits the paper’s impact. The proposed method is a projection of branching diffusion’s WFR flow and does not indicate what conservation would require for the other set-level generators the paper itself cites (PSD, EventFlow, Edit Flows), and it applies only when the counts are exact, known in advance, and defined on grid-aligned regions, while for example budgets given as ranges are deferred to future work.

Missing baselines The paper never compares against a count-conditionally trained model and only against soft conditioning.

Limited method The method is limited to grid cells and can only approximately handle the application subset that would be round etc.. This for example leads to situations, where the abstract and Figure 1 claim all 33 counts are met, but the constraint actually satisfied is not the one the application specifies.

Theory breaks for overlap In Appendix B they show that the derived method does not work for the overlapping course and resolves to other measures. A simple positional diffusion model could have done the same here, given that the table is attained not from the method.

The learned component might not contribute as implied CMMD is close to Uniform-Budget (.257 vs .276 on Syn-1, .289 vs .329 on Earthquake), while MMD shows a 5.7× gap on the same setting. The authors state that the gain lies in how events are placed within cells without showing it this, and never compare to values for a table drawn from w_ref with uniform within-cell placement.

Questions for authors

see weaknesses

3 comments
AuthorsOpenReview ↗

Response to Reviewer 5WA7 (1/2)

Thank you for your thorough review and constructive feedback. We address your concerns below and hope these clarifications will help you re-evaluate and update the score:

Summary

  • Scope (point 1). One network trained without budgets serves every budget on the grid, and we now separate theory from implementation.
  • New controls (points 2 and 5). We add a count-conditioned baseline (Table S2) and a fitted-table control that isolates learned placement (Table S1).
  • Grid (point 3). The wording now says exactness holds on the gridded areas, and Table 3 reports errors on continuous disks.
  • Theory and the position model (point 4). The table and position parts are modular by design. The theory shows why this split is forced and specifies the position dynamics. Appendix B's example is a counterexample to the projected intensity flow, not our sampler.

1. Scope and impact

"The motivating application is very narrow ... combined with the scope of the method it substantially limits the paper's impact."

We agree that impact depends on reach, and the method's key property is to train once without budgets, sample under any budget on the grid. On a fixed grid, Eq. (3) separates the budget-dependent table law from the position law given the table, and every region is a union of cells, so one position network serves every budget.

Wildfire coverage is one instance, and related problems arise in network tomography, fixed-margin tables and spatial disaggregation [1–3]. Below we separate what the theory covers from what our implementation covers.

  • Other generators. Proposition 1(ii) holds for any pure-jump process with bounded rates and gives an explicit test. Counts stay exact on every sample if and only if every jump keeps the cell table or changes it by an integer null vector of HH. Insertions and deletions, as in PSD or edit-based flows, would be fused into such moves. Transport, as in EventFlow, may cross a cell boundary only between cells in the same regions. Building such dynamics is generator-specific work, and our method uses only updates that pass this test.
  • Counts known in advance. We agree that counts are random in the data. The data law is a mixture of count-conditional laws, and an observed or planned budget selects one component.
  • Exact counts, not ranges. A range is a union of exact budgets, so by Eq. (3) its target is a mixture of exact-budget targets. Only the table sampler changes, but we have not run this, so it remains future work.
  • Grid alignment. The theory needs only cells of which every region is a union (Section 3.2), not a grid. Any finite set of regions cuts the domain into smallest pieces, and these pieces can serve as cells. We use grid boxes because reflected transport and uniform refresh then have closed forms.

Exactness also matters downstream, as a new experiment on 156 test weeks shows (Appendix E.4). Counts right only on average overstate unserved ignitions in every week, by 44% when daily totals are exact.

2. Count-conditionally trained baseline

"The paper never compares against a count-conditionally trained model and only against soft conditioning."

We agree and have added it (Table S2, Appendix E.2): count conditioning improves the repaired baseline, but ours remains lower on all three metrics. We trained branching diffusion with the budget bb as input in training and sampling, with matched backbone width, diffusion steps and epochs.

Before repair, 79–100% of its samples violate at least one count. The mechanism is that of Proposition 1(i), since independent births and deaths let each sample's counts drift even with the budget as input. Entries are geometric-mean ratios to Uniform-Budget over the 16 partition settings, as in Table 1 (lower is better).

MethodMMDCMMDSTSD
BD + Repair0.910.930.83
Count-conditioned BD + Repair0.740.780.80
Ours0.140.290.71

3. Grid approximation

"the constraint actually satisfied is not the one the application specifies"

We agree the wording was imprecise, and Table 3 already measures error on the continuous disks the application specifies. With the same trained network, constrained sampling gives disk and intersection errors of 2.55 and 2.69, against 3.40 and 3.30 under projection afterwards. Exactness holds on the grid-cell unions that approximate the circular areas, and the abstract and Figure 1 now state this.

Exact counts on the disks themselves are possible in theory, but the curved-cell sampler is engineering beyond this paper. Gridding is also a fixed redefinition, the same in every sample, whereas count errors vary between samples and overstate unserved ignitions by 44% in the experiment above.

AuthorsOpenReview ↗

Response to Reviewer 5WA7 (2/2)

4. Theory and sampling under overlap

"Theory breaks for overlap ... A simple positional diffusion model could have done the same here, given that the table is attained not from the method"

You are right that the table and the positions are sampled by separate parts, and this is by design. The chain rule over the table (Eq. (7)) splits the target into a table law and a position law given the table, and the table-first sampler makes the split explicit. We should have stated more clearly what the theory contributes to each part.

(i) The theory shows the split is forced

Section 4.1 and Appendix B study the natural extension of branching diffusion, projecting its intensity flow so that expected counts stay fixed. This fails in two ways. Independent births and deaths let each sample's counts drift (Proposition 1(i)). Under overlap, the projected flow also settles at the wrong allocation: in Appendix B's three-cell example it puts mass 0.618 on each outer cell, where the conditioned reference puts 1/2.

Proposition 1(ii) then shows that any sampler with exact counts must change the table by integer null-space steps. Our sampler is built from such steps, and it does not run the projected flow.

(ii) At a fixed table, the theory specifies the position dynamics

Restricting the WFR flow to a fixed table gives transport reflected at the cell walls plus uniform refresh within each cell. The refresh matches the constrained growth to first order near the cell-uniform intensity (Appendix C). The chain ends exactly at uniform positions in the cells (Lemma 2), and each reverse step has an explicit per-event posterior (Appendix D.2), so no birth–death rate network is learned.

(iii) The table moves matter in practice

The Metropolis–Hastings chain targets the fitted table law and reaches every feasible table (Proposition 4). The network's prediction at k=Kk=K sets its proposal weights, which changes the chain's speed but not its target (Appendix D.4). Disabling the table moves raises MMD from 0.067 ± 0.003 to 0.081 ± 0.004 on the weekly wildfire budgets, with both variants meeting every gridded count.

5. Contribution of learned positions

"never compare to values for a table drawn from w_ref with uniform within-cell placement"

We have added this control (Table S1, Appendix E.2). On the same tables sampled from the fitted table law, learned instead of uniform placement lowers MMD on every dataset (Total budgets, mean ± s.d. over five runs):

DatasetUniform placementLearned placement
SYN-10.142 ± 0.0030.033 ± 0.002
SYN-20.124 ± 0.0030.031 ± 0.002
EARTHQUAKE0.336 ± 0.0040.213 ± 0.008
EMS0.466 ± 0.0020.054 ± 0.003

We agree that the small CMMD gap needs an explanation. Under Joint budgets all feasible methods share the same cell counts, so differences from Uniform-Budget arise from within-cell placement (Uniform-Budget → Ours):

MetricSYN-1EARTHQUAKE
MMD0.131 → 0.0230.318 → 0.108
CMMD0.276 → 0.2570.329 → 0.289

There are 512 cells but at most 158 events per window in these two datasets (19–67 on the synthetic data, 50–158 on EARTHQUAKE, Appendix E.5). With fewer than one event per three cells, many occupied cells hold one or two events, and their within-cell histograms have little to compare. MMD instead pools whole-window features such as the gaps between consecutive events, which placement changes directly (Appendix E.1).

References

[1] Tebaldi & West. Bayesian inference on network traffic using link count data. JASA, 1998.

[2] Diaconis & Sturmfels. Algebraic algorithms for sampling from conditional distributions. Annals of Statistics, 1998.

[3] Taylor, Andrade-Pacheco & Sturrock. Continuous inference for aggregated point process data. JRSS-A, 2018.

I thank the authors for the additional experiments that adresses some of my points.

I still appreciate the method, however, my main problem remains: This paper tackles the niche area of point processes and count constrained point processes are even more niche within that. Yet, it gets even more niche/constrained by only tackling case where the count is constrained on a regular grid with exact counts. To attain that the model is specifically designed for this type of inference (fixed grid, cell-preserving dynamics) at training time and does not naturally generalize to further generators and conditioning. All of this erodes the significance and contribution, consequently I can not increase my score.

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

The paper propose Count-Conserving Diffusion on event data generation for planning task. The method decomposes generation into sampling a cell-count table and sampling event locations conditioned on that table. A Poisson mixture model is first fitted to cell-count tables extracted from the training data and then conditioned on the prescribed count budget. A denoising network, together with the corresponding reverse sampling procedure, generates event configurations that approximate the true conditional distribution. Within-cell transport and refresh operations preserve the cell-count table throughout position denoising. For overlapping regional constraints, the authors introduce null-space updates that explore different feasible cell-count allocations while preserving all prescribed regional counts. These updates incorporate Metropolis–Hastings acceptance rules and Markov-basis moves to ensure convergence to the fitted conditional table distribution under the stated assumptions. Experiments show that the method satisfies the regional count constraints while producing event configurations that more closely match the real data distribution than competing methods that also satisfy the constraints exactly.

Critical strengths
  1. Clear methodological contribution. The paper proposes Count-Conserving Diffusion, combining explicit count-conservation constraints with diffusion-based generation, enabling samples to satisfy prescribed regional counts while capturing the structure of the underlying data distribution.
  2. Clear presentation. The problem formulation and main contributions are clearly articulated. The paper is well organized, and the progression from the theoretical motivation to the proposed method is easy to follow.
  3. Good experimental evaluation. The experiments include real-world datasets, which strengthens the practical relevance of the evaluation. They assess both count-constraint satisfaction and generation quality, providing empirical support for the proposed approach.
Critical weaknesses
  1. The practical motivation for enforcing exact count conservation remains insufficiently justified. While the authors argue that scenarios with different event counts correspond to different planning problems, it is unclear whether small deviations from the prescribed counts would meaningfully affect downstream planning decisions. Moreover, in applications such as wildfire management, event counts are inherently stochastic rather than determined by the locations or coverage areas of service bases. Without such clarification, the proposed method risks becoming merely a technical showcase with limited real-world relevance. 2.The paper lacks direct ablation studies to quantify the contribution of each proposed component and demonstrate how much each component improves the overall performance.
Questions for authors
  1. Please see Weakness One. I would like to understand how significantly violations of count conservation affect the training of downstream planning tasks. Why is it necessary for the generated data to strictly satisfy the count constraints? Can generated data that does not fully satisfy these constraints still be useful for training planning models? If authors can clarify the basic motivation of strict count-conserving, I would be willing to reconsider my rating.
  2. I am curious about the contribution of each proposed component to the final performance. The authors are encouraged to provide direct ablation studies on individual components, including the introduction of the count table and the null-space updates, to quantify their respective contributions and demonstrate their effectiveness.
4 comments
AuthorsOpenReview ↗

Response to Reviewer PECu (1/2)

Thank you for your thorough review and constructive feedback. We address your concerns below and hope these clarifications will help you re-evaluate and update the score:

W1/Q1: Why exact count conservation?

"it is unclear whether small deviations from the prescribed counts would meaningfully affect downstream planning decisions"

Yes, they do. In a new experiment on 156 real test weeks, counts right only on average overstate unserved ignitions by 44% when daily totals are exact, and by 52% under Poisson scatter (Appendix E.4). They also enlarge the fleet needed for 95% service by 16–19%. Counts are random in the data, but planning asks for the conditional law given the budget.

We agree that ignitions are random and are not caused by where bases are. The coverage areas only define the regions in which a workload is counted, such as 30 ignitions near base A and 21 near base B in one week. In our wildfire evaluation these counts are each held-out week's observed counts (Section 5.3), and a planner can also set them as a stress test. We then generate the events behind the counts, Pdata(N∣Φ(N)=b)P_{\mathrm{data}}(N\mid\Phi(N)=b).

When do approximate counts suffice for training?

"Can generated data that does not fully satisfy these constraints still be useful for training planning models?"

Yes, if the downstream objective depends only on average load. Capacity and peak-driven planning instead has a convex overflow loss ∑(load−capacity)+\sum(\text{load}-\text{capacity})_+, because an overloaded base loses ignitions while quiet-day spare capacity cannot be banked. A plan trained or calibrated on scattered counts therefore absorbs fake peaks and over-provisions.

Suppose a base's real week has 10 ignitions on its busiest day and the base can handle 10. Scenarios giving 8 or 12 are right on average, but every 12 is a fake overload that the plan buys capacity for.

New planning experiment

To isolate count errors from placement errors, each scenario reuses the real week's ignitions. We keep, drop or duplicate each ignition at random, or redraw each day's ignitions from that day with the daily total fixed. Every event therefore sits at a real time and place, and every area count is right on average but varies around the real one (Appendix E.4).

ScenarioCount variance / budgetUnserved ignitions vs exactFalse alarms / missed per weekFleet for 95% service
Exact (the real week)0+0%0 / 0254 units
Right on average, daily totals exact0.81+44%10.4 / 4.4+16%
Right on average, Poisson0.99+52%11.2 / 4.7+19%
Right on average, twice Poisson1.99+99%14.9 / 5.9+41%

The real weeks average 16.5 truly overloaded base-days, and the scattered scenarios overstate unserved ignitions in all 156 weeks. The Poisson row is also a floor for any Poisson or Cox generator that draws events from an intensity Λ\Lambda, because a random intensity only adds scatter:

Var⁡N(R)=EΛ(R)+Var⁡Λ(R)≥EN(R)for every region R.\operatorname{Var} N(R) = \mathbb{E} \Lambda(R) + \operatorname{Var} \Lambda(R) \ge \mathbb{E} N(R) \quad \text{for every region } R.

Fixing counts afterwards

Under overlap, repairing counts after sampling also costs fidelity, raising MMD from 0.067 to 0.117 with our network fixed (W2/Q2 table, second comment).

Related problems

Traffic analysis faces the same problem, sampling the journeys that reproduce all road-counter totals although one journey passes several counters [1]. Fixed-margin tables [2] and spatial disaggregation [3] condition on totals in the same way, and recent diffusion models keep given measurements or constraints exactly [4–6]. Our contribution is this setting for event sets with overlapping integer counts, where exact count conditions are required.

AuthorsOpenReview ↗

Response to Reviewer PECu (2/2)

W2/Q2. Contributions of the components

"The paper lacks direct ablation studies to quantify the contribution of each proposed component"

We agree and have added controls for the fitted table distribution, learned within-cell positions and null-space updates, and we compare constrained sampling with post-hoc projection using the same network. Every component lowers MMD (Appendix E.2 and E.3), and keeping counts during sampling lowers it from 0.117 to 0.067.

Count table and within-cell positions

"Fitted table" draws tables from wrefw_{\mathrm{ref}}, our fitted Poisson-mixture model of cell counts, and places events uniformly within cells. "Full method" adds learned placement on the same tables. We report MMD under Total budgets (mean ± s.d. over five runs):

DatasetUniform-BudgetFitted tableFull method
SYN-10.349 ± 0.0020.142 ± 0.0030.033 ± 0.002
SYN-20.343 ± 0.0030.124 ± 0.0030.031 ± 0.002
EARTHQUAKE0.659 ± 0.0010.336 ± 0.0040.213 ± 0.008
EMS0.787 ± 0.0010.466 ± 0.0020.054 ± 0.003

The fitted table lowers MMD by 41–64% over Uniform-Budget, and learned placement lowers it by a further 37–88% on the same tables (Table S1, Appendix E.2).

Null-space updates under overlap

Disabling null-space updates on overlapping wildfire budgets raises MMD from 0.067 ± 0.003 to 0.081 ± 0.004, about 21%, with each sample's initialization and trained position network unchanged. Both variants satisfy all prescribed grid counts. Under overlap the same counts can be met by different tables, and null-space updates move between them while keeping every count.

With region A = cells {1, 2} and region B = cells {2, 3}, one event in each can sit in cell 2 or in cells 1 and 3.

Constrained sampling vs post-hoc projection

With the same trained network, keeping counts during sampling improves all four metrics over sampling under daily totals and then projecting onto the coverage budget (Table 3, Appendix E.3). Under overlap one ignition can lie in up to six coverage areas, so correcting one count disturbs others. Global projection of DSTPP samples accordingly rewrites a median of 68–84% of the output events (Section 5.3).

Error (lower is better)Post-hoc projectionCounts kept during sampling
MMD0.1170.067
Ignitions within 90 km of a base3.402.55
Ignitions in areas reached by two bases3.302.69
Ignitions on a base's busiest day1.731.63

We include these ablations in the revised Sections 1 and 5 and Appendix E.

References

[1] Tebaldi & West. Bayesian inference on network traffic using link count data. JASA, 1998.

[2] Diaconis & Sturmfels. Algebraic algorithms for sampling from conditional distributions. Annals of Statistics, 1998.

[3] Taylor, Andrade-Pacheco & Sturrock. Continuous inference for aggregated point process data. JRSS-A, 2018.

[4] Wang, Yu & Zhang. Zero-shot image restoration using denoising diffusion null-space model. ICLR, 2023.

[5] Lou & Ermon. Reflected diffusion models. ICML, 2023.

[6] Cardei et al. Constrained discrete diffusion. NeurIPS, 2025.

Dear Reviewer PECu,

Thank you again for your review. Before the discussion period closes, we wanted to briefly check whether our revision addressed your two main concerns.

Why exact counts matter. Your question on this was a fair one, and the submission did not answer it clearly. The setting itself has a long history. Traffic analysis samples the journeys that reproduce all road-counter totals (Tebaldi & West, 1998), and statistics samples tables with fixed margins (Diaconis & Sturmfels, 1998). Like these methods, our model takes the observed counts as given and generates the events behind them.

Exactness matters for planning because of peaks. If a base can handle 10 ignitions on its busiest day, scenarios giving 8 or 12 are right on average, yet every 12 is a false overload that the plan pays for. On 156 real wildfire weeks, such counts overstate unserved ignitions by 44% (Appendix E.4).

Ablations. We also added the ablations you asked for in your second question, covering the count table, learned placement and null-space updates (Appendices E.2–E.3).

Please refer to our detailed response above for the full results. We would be glad to hear whether these address your concerns, and we welcome any further questions.

Best regards,

The Authors

Thank you for your response. However, the two historical examples you cite do not appear to share the same motivation.

Regarding the wildfire example, emergency-response planning should ensure that available capacity can accommodate demand with sufficiently high reliability. Historical records characterize a distribution of event counts within a region but not a known exact count. Capturing uncertainty and variability of counts therefore seems more important than exact ones. It's still unclear that why exact count does matter.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response to all reviewers and the AC

We appreciate the time and care the reviewers and the AC have given this submission. Beyond the strengths the reviewers noted, a rigorous formulation and strong experiments, their reviews identified four gaps that the submission left open, and we agree with each of them:

  • The paper did not make clear why planning needs exact rather than approximately right counts (PECu, 5WA7).
  • The interleaved sampler had no stated joint target (QQTb).
  • Two comparisons were missing, a count-conditioned baseline and a control that isolates learned placement (5WA7, PECu).
  • The abstract and Figure 1 overstated exactness on the circular coverage areas (5WA7).

The revision addresses each gap, mostly with new experiments, and new or changed text in the revised PDF is in blue.

Summary of changes

  1. Why exact counts matter for planning (PECu, 5WA7). A new experiment on 156 real wildfire test weeks (Appendix E.4) shows that scenarios whose counts are right only on average overstate unserved ignitions in every week, by 44% when daily totals are exact, and enlarge the fleet needed for 95% service by 16–19%.
  2. Table-first sampler (QQTb). Following QQTb's suggestion, Algorithm 1 now completes the table chain before denoising positions, so the position sampler no longer depends on the table's path. At an equal proposal budget the results match the earlier interleaved schedule (Table S3).
  3. New baselines and ablations (PECu, 5WA7). A count-conditioned branching-diffusion baseline (Table S2), a fitted-table control with uniform placement that isolates learned placement (Table S1), and a null-space ablation (Appendix E.3).
  4. Scope and wording (5WA7). The abstract and Figure 1 now say that exactness holds on the gridded coverage areas, and Section 6 states the limits on region shape and range budgets.

The replies to each reviewer give the details.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.