acceptodds
Under review as a conference paper at ICLR 2027

Sinkhorn-Reparameterized Primal-Dual Optimization for Scalable Quadratic Assignment

Abstract

Finding high-quality solutions to large-scale instances of the quadratic assignment problem (QAP) remains challenging under limited computational budgets. We approach QAP through an equivalent continuous formulation with entrywise integrality constraints over the doubly stochastic domain. Building on this formulation, we propose Sinkhorn-Reparameterized Primal-Dual Optimization (SRPD), which replaces explicit primal projection with a differentiable Sinkhorn parameterization. A finite sequence of Sinkhorn normalizations enables gradient-based primal search in an unconstrained latent space, while entrywise dual updates adapt the objective in response to integrality violations. SRPD then recovers feasible permutations through trajectory harvesting: search trajectories run in parallel on GPUs, and their soft assignments are greedily decoded throughout optimization to retain the best solution. For the finite Sinkhorn map and greedy decoder, we bound the decoding error and returned objective using integrality violations and soft-objective values, and identify sufficient conditions for exact penalization. Experiments on QAPLIB demonstrate competitive solution quality and runtime, with substantial improvements over the projected primal–dual baseline, reducing the average gap from 19.28% to 0.50%. On larger taiXXeYY instances, SRPD achieves substantially lower final gaps than the compared baselines under a common time budget.

Primary area:
optimization
Keywords:
Quadratic Assignment Problems, Sinkhorn Algorithm, Primal-Dual

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_hYauOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper aims to solve Quadratic Assignment Problem (QAP) which is fundamental and extremely difficult field in combinatorial optimization. The author proposed Sinkhorn-Reparameterized Primal-Dual Optimization (SRPD), a scalable framework for solving QAP. There are major two contributions:

  1. The author reformulates QAP as a continuous optimization problem with integrality constraints and replace explicit projection with differential Sinkhorn operation so that enables unconstrained latent-space optimization with adaptive primal–dual updates.

  2. Building on this formulation, the paper develops a GPU-parallel trajectory harvesting strategy that greedily decodes intermediate soft assignments across multiple optimization runs and retains the best feasible permutation, exploiting candidate solutions encountered throughout the optimization process.

Extensive experiments on QAPLIB, large-scale taiXXeYY instances, and graph isomorphism benchmarks establish the framework’s empirical effectiveness

Critical strengths
  1. The paper provides a clear formulation of the QAP and develops a coherent optimization framework. The proposed Sinkhorn reparameterization offers an intuitive, differentiable alternative to explicit projection, enabling optimization in an unconstrained latent space. Its integration with primal–dual updates and trajectory harvesting further strengthens the framework’s computational practicality.

  2. The paper presents a detailed theoretical analysis of the proposed framework, covering the properties of Sinkhorn reparameterization, bounds on greedy decoding errors, and sufficient conditions for exact penalization. Notably, the analysis explicitly accounts for a finite number of Sinkhorn iterations rather than assuming exact normalization, bringing the theoretical treatment into closer alignment with the practical implementation.

  3. The experiments span standard QAPLIB benchmarks, large-scale taiXXeYY instances, and graph isomorphism problems. The proposed method substantially outperforms the projected primal–dual baseline in both solution quality and runtime. Results on larger instances further demonstrate its practical scalability, while optional 2-opt refinement yields additional gains in solution quality.

Critical weaknesses
  1. Although the paper derives useful bounds for decoding errors and objective values, it remains unclear how these results support the adaptive optimization scheme used in practice. The exact-penalty analysis in Theorem 4.2 assumes fixed penalty coefficients of sufficient magnitude, yet the algorithm updates its dual variables adaptively without guaranteeing that the resulting values satisfy those assumptions. This mismatch is particularly evident in the experiments on larger taiXXeYY instances, where the dual variables are kept positive—a configuration outside the stated sufficient conditions. A further gap is the absence of a convergence analysis for the adaptive primal–dual procedure as a whole. The authors should therefore explain more precisely what the theoretical results imply for the implemented algorithm and where their applicability ends.

  2. The substantial gains over the projected primal–dual baseline demonstrate the effectiveness of the overall framework, but leave the sources of those gains unresolved. The current evaluation provides limited evidence for distinguishing the impact of Sinkhorn reparameterization from the benefits of parallel multistart optimization and trajectory harvesting. A more informative assessment would vary these components in controlled ablation studies while holding the computational budget constant, allowing their respective contributions to be assessed. In addition, the stochastic optimization process calls for an evaluation across independent runs. Reporting the resulting variability alongside aggregate performance would provide stronger support for the reliability of the experimental findings.

Questions for authors

My main concern is the alignment between the theory and implementation.

3 comments

Q1. My main concern is the alignment between the theory and implementation.

W1. ...The authors should therefore explain more precisely what the theoretical results imply for the implemented algorithm and where their applicability ends.

Thank you for highlighting the distinction between the theoretical guarantees and the implemented adaptive algorithm. We clarify the scope and applicability of each result below.

  • Theorem 4.1 (Finite-Sinkhorn recovery) bounds the assignment residual and greedy decoding error in terms of integrality violation. Eq. (8) (Harvested-output guarantee) further bounds the objective of the best retained permutation using the soft-objective values and integrality violations encountered during search. Both results apply directly to our implementation, including finite Sinkhorn normalization and adaptive dual updates, without requiring convergence or restrictions on the signs of dual coefficients.
  • Theorem 4.2 (Exact penalization and decoded accuracy) establishes that sufficiently strong fixed penalty coefficients make the latent objective's infimum equal to the QAP optimum, and that a near-optimal latent solution yields a decoded permutation with a correspondingly bounded optimality gap. This guarantee also applies to an individual adaptive iterate when its conditions are satisfied, but does not imply that the adaptive updates will reach this regime. In particular, the taiXXeYY experiments operate outside this sufficient-penalty regime.

In summary, our theory provides recovery and objective guarantees for the actual finite-budget outputs, rather than convergence guarantees for the adaptive optimization trajectory. These results do not guarantee that the search encounters a near-optimal candidate; the effectiveness of the adaptive search itself is supported empirically.

W2-1. Controlled ablation of individual components.

Thank you for this suggestion.

  • The PDBO–SRPD comparison in Table 1 of the original manuscript already provides a controlled ablation that isolates the contribution of Sinkhorn-based Birkhoff reparameterization from the other components of the framework. Specifically, both methods share the same entrywise dual updates, multi-start optimization, trajectory harvesting, and greedy decoding, differing only in their assignment constraint-handling mechanisms. We have now extended this comparison to taiXXeYY in Section 5.3.
  • The results demonstrate the effectiveness of Sinkhorn reparameterization. Replacing projection-based primal updates with Sinkhorn reparameterization reduces the family-averaged gap on QAPLIB from 19.28% to 0.50%, and the average gap on taiXXeYY from 79.36% to 4.85%, with shorter recorded times to the best solutions on both benchmarks.
MethodAssignment constraint handlingMulti-startTrajectory harvesting & greeding decodingQAPLIB Avg. GapQAPLIB Avg. Time (sec)taiXXeYY Avg. GaptaiXXeYY Avg. Time (sec)
PDBOEuclidean projection as in Eq (4)✓\checkmark✓\checkmark19.28%0.9379.36%254.17
SRPDSinkhorn parameterization✓\checkmark✓\checkmark0.50%0.414.85%234.83

We have further clarified this comparison in Sections 5.1 and 5.3 of the revised manuscript.

W2-2. Reporting the resulting variability alongside aggregate performance ...

Thank you for highlighting the importance of evaluating variability across independent runs.

  • We conducted additional experiments with five independent random seeds for PDBO and SRPD on QAPLIB. The table below reports the mean ± standard deviation of the relative gap to the BKS (%) across five runs for each instance family.
  • The results demonstrate that SRPD's substantial advantage over projection-based PDBO remains consistent across independent runs.
FamilyPDBOSRPD
bur(2.5 ± 0.12)%(0.0 ± 0.00)%
chr(76.0 ± 1.81)%(3.4 ± 0.58)%
els(38.1 ± 0.00)%(2.1 ± 1.08)%
esc(11.5 ± 0.40)%(0.0 ± 0.01)%
had(3.9 ± 0.33)%(0.0 ± 0.00)%
kra(16.0 ± 0.33)%(0.1 ± 0.11)%
lip(1.2 ± 0.02)%(0.3 ± 0.02)%
nug(8.4 ± 0.15)%(0.0 ± 0.01)%
rou(13.4 ± 0.54)%(0.1 ± 0.05)%
scr(21.7 ± 1.55)%(0.0 ± 0.01)%
sko(11.4 ± 0.13)%(0.7 ± 0.02)%
ste(42.3 ± 1.90)%(0.4 ± 0.19)%
tai(11.7 ± 0.14)%(1.1 ± 0.06)%
tho(13.8 ± 0.97)%(0.6 ± 0.06)%
wil(7.2 ± 0.20)%(0.4 ± 0.02)%
Avg. Gap(18.60 ± 0.19)%(0.62 ± 0.09)%

We have included the complete five-seed results in Appendix F of the revised manuscript.

Reviewer_hYauOpenReview ↗

Reply to rebuttal

I have no further concerns, the author have addressed my questions.

Thank you for your positive feedback. We are glad that our response has addressed your concerns.

Reviewer_yPJ6OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The authors propose Sinkhorn-Reparameterized Primal-Dual Optimization (SRPD), a method to solve the quadratic assignment problem by replacing primal projection with a differentiable Sinkhorn parameterization. Using a finite sequence of Sinkhorn normalizations, it optimizes the latent representation with primal-dual updates and entrywise integrality penalties, while running parallel trajectories and frequently decoding them greedily to retain the best feasible permutation. The authors add a theoretical analysis by bounding the decoding error and returned objectives. Experiments on QAPLIB and larger taiXXeYY benchmark instances show promising improvement compared to the projected primal-dual baseline.

Critical strengths
  1. The paper combines Sinkhorn parameterization with primal-dual optimization practically and coherently. The experiments show a clear improvement over the baseline.
  2. The authors analyze the effects of using a finite number of Sinkhorn iterations. This adds useful theoretical insight.
Critical weaknesses
  1. Limited novelty. SRPD combines known/established techniques and doesn't demonstrate enough how much of the improvement comes specifically from the Sinkhorn parameterization rather than from the other components (such as trajectory harvesting and greedy decoding). This makes it difficult to assess the paper's main contribution.
Questions for authors
  1. Could you provide an ablation study that isolates the contribution of the Sinkhorn parameterization from other components? This would help clarify which aspects of SRPD are responsible for the improvements over PDBO and improve your contribution claim.
3 comments

Q1. Could you provide an ablation study that isolates the contribution of the Sinkhorn parameterization from other components? This would help clarify which aspects of SRPD are responsible for the improvements over PDBO and improve your contribution claim.

Thank you for this constructive suggestion.

  • Actually, our PDBO–SRPD comparison in Table 1 of the original manuscript has already provided a controlled ablation that isolates the contribution of Sinkhorn-based birkhoff reparameterization from the other components of the framework. Specifically, both PDBO and SRPD retain the same entrywise dual updates, multi-start optimization, trajectory harvesting, and greedy decoding, differing only in their primal constraint-handling mechanism. This comparison shows that replacing projection-based primal updates with Sinkhorn reparameterization reduces the family-averaged gap on QAPLIB from 19.28% to 0.50% with shorter computational time.
  • To further show the contribution of sinkhorn parameterizaion in SRPD, we additionally replaced the sinkhorn map with the Birkhoff stick-breaking parameterization proposed by [1], as suggested by Reviewer Lch5. We retained SRPD's other components and tested the performance of the three variants on both QAPLIB and taiXXeYY instances. The configurations and solution-quality results are summarized below:
MethodAssignment constraint handlingMulti-startTrajectory harvesting & greeding decodingQAPLIB Avg. GapQAPLIB Avg. Time (sec)taiXXeYY Avg. GaptaiXXeYY Avg. Time (sec)
PDBOEuclidean projection as in Eq (4)✓\checkmark✓\checkmark19.28%0.9379.36%254.17
Stick-BreakingStick-breaking parameterization[1]✓\checkmark✓\checkmark0.77%8.4819.97%601.83
SRPDSinkhorn parameterization✓\checkmark✓\checkmark0.50%0.414.85%234.83

Overall, these controlled comparisons support two complementary conclusions:

  • First, replacing projection-based primal updates with differentiable reparameterization contributes substantially to the effectiveness of the framework, while holding other components fixed.
  • Second, the specific choice of parameterization also matters, with Sinkhorn providing a clear additional advantage on the larger structured instances evaluated in this work.

We have clarified this in Section 5.1 and added detailed ablation results in Section 5.3 and Appendix E of the revised manuscript.

[1] Linderman, Scott, et al. "Reparameterizing the birkhoff polytope for variational permutation inference." International Conference on Artificial Intelligence and Statistics. PMLR, 2018.

Reviewer_yPJ6OpenReview ↗

Reply to Rebuttal

Thank you for the response and for providing the additional results. They address my concern.

Thank you for your positive feedback. We are glad that our response has addressed your concerns.

Reviewer_Lch5OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors focus on approximate solution of QAP, which is the problem (1) over permutation matrices. For that, they adapt the method by (Liu et al, 2025), which represent the integrality constraints by equality constraints g(X_{ij})=0, where g can be, eg, a suitable quadratic function. The method dualizes these g-constraints and use gradient descent-ascent on the Lagrange function. Without the g-constraints, teh problem minimizes the QAPobjective over the set BnB_n of doubly-stochastic matrices. The dual updates in fact monotonically change to progressively increase the integrality penalty and make the problem more and more non-convex.

In the direct implementation (4) of this method, the primal updates have to project onto BnB_n. The innovation of the paper is to parameterize the relative interior of BnB_n by the Sinkhorn map S:Rn×n→ri(Bn)S: \mathbb{R}^{n\times n}\to\mathrm{ri}(B_n), which is evaluated in the usual way by first exponentiating, and then alternating normalization of rows and columns (aka iterative proportional fitting, IPF). This converts the relaxed QAP to unconstrained one. In reality, the infinite number of these normalizations is truncated to a finite number KK, and the map's SKS_K derivative is then computed by backprop. This basic algorithm is then finished by applying a primal heuristic (to round XX to a perumtation matrix) and by multiple initializations.

The experiments compare the proposed method to several SOT methods for approximating QAP, and also to the direct method (Liu et al, 2025), given by (4), which uses projections to BnB_n. This is done on public QAP instance datasets QAPLIB (classical small-size instances) and taiXXeYY (larger instances). On QAPLIB, the method is a slight piece worse than SOAT methods. On taiXXeYY, it mostly beats all SOTA methods.

Critical strengths

The method is simple and comparable to SOTA methods on QAPLIB, but convincingly better than SOTA on taiXXeYY.

Critical weaknesses

I see no major technical issues and I like the paper in general. However, I am not totally sure to how much real novelty it brings.

The method is firmly rooted in (Liu et al, 2025), which is applied to QAP with the doubly-stochastic constraint replaced by the truncated Sinkhorn map. Calling it a primal-dual method is somewhat misleading: the QAP constraint X∈PnX\in P_n (or X∈BnX\in B_n) is not dualized, dualized are only the added integrality constraints. Alternatively, this could be called a primal gradient descent augmented with continuation / temperature change. Note that it is not shown that the relaxed problem is convex for any value of penalty β\beta, so the continuation does not start with a convex function. As such, the method inherits some drawbacks of purely primal local search: if the objective landscape has many valleys separated by barriers with high objective values, the iterates can remain in a single valley forever. Multiple initializations need not help, since these valleys can be exponentialy many.

For QAP in particular, I would expect problems for instances where matrices F,DF,D have entries of very different values. In extreme, if F,DF,D have only values 0 or +∞+\infty, finding a feasible (= with finite QAP objective) permutation PP is NP-complete (reduction from graph isomorphism). I believe this is the case also for the relaxed QAP, where PnP_n is replaced with BnB_n. While dual method can get good lower bounds even in this case, primal local search can be very poor; in fact finding a good initial XX (or ZZ) can be infeasible. That said, the method empirically performs very well on instances of graph isomorphism problem, where F,DF,D have 0-1 entries (in the appendix).

At the first sight, using the Sinkhorn map may look like a main innovation. However, parameterizing the feasible set of a constraint problem to obtain an unconstrained problem is common in optimization. For Birkhoff polytope, the Sinkhorn map is one such well-known parameterization. But there are other parameterizations, eg, the stick-breaking transformation by [Linderman et al, AISTATS 2018]. One can hypothesize that using this parameterization would lead to comparable empirical results. True, an obstacle may be that the stick-breaking tx is only piecewise smooth.

Note, moreover, that parameterization of the feasible set could be replaced by adding a barrier function to the objective. Eg, we could eliminate redundant variables from the linear system X1=XT1=1X1=X^T1=1 and impose nonnegativity of XX by a log-barrier. True, this would need a modification of GD to prevent jumping over the barrier. But, again, one can hypothesize that the empirical results would be comparable.

To summarize so far, I am not convinced that the empirical success is mainly due to the used Sinkhorn reparameterization: an alternative parameterization can work just as well. This is a suggestion for an ablation study (which is missing).

A bit off-topic, let me hypothesize that a suitable version of nonconvex ADMM, with enough care, might work comparably well on the tested QAP datasets.

Though the authors provide a thorough theoretical analysis, to my understanding convergence of the method is not proved. The results suggest convergence of primal updates with a fixed dual penalty, but that is changed durign the actual method.

Questions for authors

If the Sinkhorn map was replaced by a different parameterization of the Birkhoff polytope, do you believe this made a big difference? If not, can you please provide arguments?

Do you agree that for F,DF,D with large entries, the method could get stuck in poor suboptima, even with multi-init? Did you observe this failure case in your experimenting?

4 comments

W1. Calling it a primal-dual method is somewhat misleading...

Thank you for raising this point.

  • Our proposed SRPD follows a partial-Lagrangian formulation, where only a subset of constraints is dualized, while the remaining constraints are handled in the primal space [1].
  • Following Liu et al. [2], we dualize the integrality equalities; however, while Liu et al. handle simple coordinate-wise box constraints, SRPD accommodates the coupled row and column assignment constraints through Sinkhorn reparameterization.

We therefore retain the term primal–dual to describe our framework.

[1] Vandenberghe, Lieven, and Stephen Boyd. Convex optimization. Vol. 1. Cambridge: Cambridge university press, 2004.

[2] Liu, Wenbo, et al. "Smoothing Binary Optimization: A Primal-Dual Perspective." arXiv preprint arXiv:2509.21064 (2025).

W2. For QAP in particular, I would expect problems for instances where matrices...

Please see our reply to Q2.

W3. ... But there are other parameterizations, eg, the stick-breaking transformation by [Linderman et al, AISTATS 2018]. One can hypothesize that using this parameterization would lead to comparable empirical results...

Please see our reply to Q1.

W4. Barrier and ADMM methods

Thank you for suggesting these two baselines.

  • Barrier: We enforce the row and column sum constraints through the redundant affine map X=11⊤/n+HZHX=\mathbf{1}\mathbf{1}^{\top}/n+HZH, where H=I−11⊤/nH=I-\mathbf{1}\mathbf{1}^{\top}/n, and add −λt∑ijlog⁡(nXij)-\lambda_t\sum_{ij}\log(nX_{ij}). We retain SRPD’s other components.
  • ADMM: We adapt ℓp\ell_p-box ADMM [3] with p=2, objective normalization, and a binary-equivalent positive-semidefinite reformulation. We use multiple random permutation starts and periodic Hungarian decoding.

We implemented both algorithms (details are provided in Appendix G of the revised manuscript) and evaluated them on all QAPLIB and taiXXeYY instances with the same time limit of 1,800s for each instance. The table below reports the mean relative gap and average time to the best solution found on QAPLIB and taiXXeYY.

MethodQAPLIB Avg. GapQAPLIB Avg. Time (sec)taiXXeYY Avg. GaptaiXXeYY Avg. Time (sec)
Barrier4.67%11.68172.50%1139.95
ADMM14.88%8.73129.17%134.65
SRPD0.50%0.414.85%234.83
  • SRPD delivers superior solution quality on both benchmarks, achieving average gaps of just 0.50% on QAPLIB and 4.85% on taiXXeYY—over 89% lower than those of both ADMM and Barrier.

[3] Wu, Baoyuan, and Bernard Ghanem. "ℓp\ell_p-Box ADMM: A Versatile Framework for Integer Programming." IEEE transactions on pattern analysis and machine intelligence 41.7 (2018): 1695-1708.

W5. Though the authors provide a thorough theoretical analysis, to my understanding convergence of the method is not proved...

Thank you for raising this point. Our main theoretical guarantees concern the recovered discrete solutions rather than convergence of the optimization trajectory. Specifically, Theorem 4.1 and Eq. (8) bound the decoding error and the returned permutation's objective under finite Sinkhorn normalization and adaptive dual updates, without requiring convergence. Theorem 4.2 provides a complementary conditional guarantee under sufficiently strong fixed penalty coefficients. Establishing convergence of the full adaptive dynamics remains an important direction for our future work.

Q1. If the Sinkhorn map was replaced by a different parameterization of the Birkhoff polytope, do you believe this made a big difference? If not, can you please provide arguments?

Thank you for this constructive suggestion. To investigate the impact of parameterization, we replaced Sinkhorn with the Birkhoff stick-breaking parameterization of Linderman et al. [4] (denoted as Stick-Breaking), while retaining SRPD's primal–dual updates, multi-start optimization, trajectory harvesting, and greedy decoding. All variants use the same iteration budgets.

We evaluated these methods on QAPLIB and taiXXeYY. The following table summarizes the average gaps and recorded times to the best solutions.

MethodAssignment constraint handlingQAPLIB Avg. GapQAPLIB Avg. Time (sec)taiXXeYY Avg. GaptaiXXeYY Avg. Time (sec)
PDBOEuclidean projection as in Eq (4)19.28%0.9379.36%254.17
Stick-BreakingStick-breaking parameterization[1]0.77%8.4819.97%601.83
SRPDSinkhorn parameterization0.50%0.414.85%234.83

The results reveal two key findings:

  • Reparameterization is important. Both Sinkhorn and Stick-breaking substantially outperform projection-based PDBO, suggesting that differentiable parameterization contributes significantly to the framework's effectiveness.
  • The choice of Sinkhorn parameterization also matters. SRPD achieves lower average gaps and shorter recorded times to best solutions found than Stick-Breaking on both benchmarks. A possible explanation is that Stick-breaking constructs assignments sequentially, so earlier allocations constrain the feasible intervals of later entries, while Sinkhorn employs smooth, coupled row and column normalizations that may facilitate more coordinated updates across the assignment matrix.

We have added the detailed comparison to Section 5.3 and Appendix E of the revised manuscript.

[4] Linderman, Scott, et al. "Reparameterizing the birkhoff polytope for variational permutation inference." International Conference on Artificial Intelligence and Statistics. PMLR, 2018.

Q2. Do you agree that for with large entries, the method could get stuck in poor suboptima, even with multi-init? Did you observe this failure case in your experimenting?

Thank you for raising this point.

  • We agree that QAP instances with large-scale coefficients may have many local optima. Since SRPD does not guarantee global optimality, it may get stuck in suboptimal solutions even with multiple initializations.
  • We observed a relevant case on the QAPLIB instance tai80b, where the positive entries of F range from 2 to 1,470 and those of D from 1 to 7,998. Despite multi-start optimization, SRPD returns an objective of 842,362,830, corresponding to a 2.9261% gap relative to the BKS of 818,415,043, compared with its overall family-averaged gap of 0.50% on QAPLIB.

We now clarify this point in Section 5.1 of the revised manuscript.

Reviewer_Lch5OpenReview ↗

Reply to Rebuttal

I thank the authors for their rebuttal. They have thoroughly addressed my major objections, adding two new experiments (comparison with other methods like nonconvex ADMM, comparison with another reparameterization of the Birkhoff polytope). Taking into account also the new experiments as a reply to reviewer hYau, my view of the paper is now more positive and I am not against acceptance (I raised my score).

With the additional experiments, my current view of the paper is as follows: it describes a method for a hard nonconvex problem that works better than SOTA on many public instances. Several ablation experiments (though brief) suggest that the components of the method (mainly, the Sinkhorn parameterization) are necessary for the success. There is not much useful theory explaining the empirical success (not even a convergence proof), so the paper should be seen as primarily empirical.

Given this, I find the added experiments, as well as the failure case (QAPLIB instance tai80b) a necessary part of the accepted paper. I insist that these be included into the final version (possibly in a supplement) as the authors promised, please do not forget to do this!

Thank you for your positive reassessment and constructive feedback. We appreciate your suggestions and confirm that all additional experiments and discussion on the failure case will be included in the final version.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.